Internet Multimedia Computing and

Service 9th International Conference

ICIMCS 2017 Qingdao China August 23
25 2017 Revised Selected Papers 1st
Edition Benoit Huet
Visit to download the full and correct content document:
https://textbookfull.com/product/internet-multimedia-computing-and-service-9th-intern
ational-conference-icimcs-2017-qingdao-china-august-23-25-2017-revised-selected-p
apers-1st-edition-benoit-huet/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Cloud Computing and Service Science 7th International

Conference CLOSER 2017 Porto Portugal April 24 26 2017
Revised Selected Papers Donald Ferguson

https://textbookfull.com/product/cloud-computing-and-service-
science-7th-international-conference-closer-2017-porto-portugal-
april-24-26-2017-revised-selected-papers-donald-ferguson/

Knowledge Graph and Semantic Computing Language

Knowledge and Intelligence Second China Conference CCKS
2017 Chengdu China August 26 29 2017 Revised Selected
Papers 1st Edition Juanzi Li
https://textbookfull.com/product/knowledge-graph-and-semantic-
computing-language-knowledge-and-intelligence-second-china-
conference-ccks-2017-chengdu-china-august-26-29-2017-revised-
selected-papers-1st-edition-juanzi-li/

Space Information Networks: Second International

Conference, SINC 2017, Yinchuan, China, August 10-11,
2017, Revised Selected Papers 1st Edition Quan Yu
(Eds.)
https://textbookfull.com/product/space-information-networks-
second-international-conference-sinc-2017-yinchuan-china-
august-10-11-2017-revised-selected-papers-1st-edition-quan-yu-
eds/

Image and Graphics 9th International Conference ICIG

2017 Shanghai China September 13 15 2017 Revised
Selected Papers Part III 1st Edition Yao Zhao

https://textbookfull.com/product/image-and-graphics-9th-
international-conference-icig-2017-shanghai-china-
september-13-15-2017-revised-selected-papers-part-iii-1st-
Foundations and Practice of Security 10th International
Symposium FPS 2017 Nancy France October 23 25 2017
Revised Selected Papers 1st Edition Abdessamad Imine

https://textbookfull.com/product/foundations-and-practice-of-
security-10th-international-symposium-fps-2017-nancy-france-
october-23-25-2017-revised-selected-papers-1st-edition-
abdessamad-imine/

Advances in Multimedia Information Processing – PCM

2017: 18th Pacific-Rim Conference on Multimedia,
Harbin, China, September 28-29, 2017, Revised Selected
Papers, Part II Bing Zeng
https://textbookfull.com/product/advances-in-multimedia-
information-processing-pcm-2017-18th-pacific-rim-conference-on-
multimedia-harbin-china-september-28-29-2017-revised-selected-
papers-part-ii-bing-zeng/

Digital TV and Wireless Multimedia Communication: 14th

International Forum, IFTC 2017, Shanghai, China,
November 8-9, 2017, Revised Selected Papers 1st Edition
Guangtao Zhai
https://textbookfull.com/product/digital-tv-and-wireless-
multimedia-communication-14th-international-forum-
iftc-2017-shanghai-china-november-8-9-2017-revised-selected-
papers-1st-edition-guangtao-zhai/

Cyberspace Safety and Security 9th International

Symposium CSS 2017 Xi an China October 23 25 2017
Proceedings 1st Edition Sheng Wen

https://textbookfull.com/product/cyberspace-safety-and-
security-9th-international-symposium-css-2017-xi-an-china-
october-23-25-2017-proceedings-1st-edition-sheng-wen/

Provable Security: 11th International Conference,

ProvSec 2017, Xi’an, China, October 23-25, 2017,
Proceedings 1st Edition Tatsuaki Okamoto

https://textbookfull.com/product/provable-security-11th-
international-conference-provsec-2017-xian-china-
october-23-25-2017-proceedings-1st-edition-tatsuaki-okamoto/
Benoit Huet
Liqiang Nie
Richang Hong (Eds.)

Communications in Computer and Information Science 819

Internet Multimedia
Computing and Service
9th International Conference, ICIMCS 2017
Qingdao, China, August 23–25, 2017
Revised Selected Papers

123
Communications
in Computer and Information Science 819
Commenced Publication in 2007
Founding and Former Series Editors:
Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu, Dominik Ślęzak,
and Xiaokang Yang

Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Phoebe Chen
La Trobe University, Melbourne, Australia
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
Junsong Yuan
Nanyang Technological University, Singapore, Singapore
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at http://www.springer.com/series/7899
Benoit Huet Liqiang Nie
•

Richang Hong (Eds.)

Internet Multimedia
Computing and Service
9th International Conference, ICIMCS 2017
Qingdao, China, August 23–25, 2017
Revised Selected Papers

123
Editors
Benoit Huet Richang Hong
Multimedia Communications Department Hefei University of Technology
EURECOM Hefei
Sophia Antipolis China
France
Liqiang Nie
Shandong University
Qingdao
China

ISSN 1865-0929 ISSN 1865-0937 (electronic)

Communications in Computer and Information Science
ISBN 978-981-10-8529-1 ISBN 978-981-10-8530-7 (eBook)
https://doi.org/10.1007/978-981-10-8530-7

Library of Congress Control Number: 2018934345

© Springer Nature Singapore Pte Ltd. 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

Printed on acid-free paper

This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
part of Springer Nature
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
 Preface

The 9th International Conference on Internet Multimedia Computing and Service

(ICIMCS 2017) was held during August 23–25, 2017, at Tsingtao, China. ICIMCS
focuses on pushing the frontier of multimedia computing technology in a variety of
application domains. During the past 8 years, ICIMCS has successfully facilitated the
communications of state-of-the-art progress made by multimedia researchers and
participants worldwide. With the venue mainly based in China, ICIMCS presents
highly-quality research papers and technical presentations organized by influential
scholars in the multimedia community. Therefore, ICIMCS offers a great opportunity
for the international community to know and get involved in the fast development of
Chinese research in multimedia such as computer vision, natural language processing,
and data knowledge understanding.
ICIMCS 2017 was organized at Shandong University of Science and Technology,
Qingdao, Shandong. We gratefully thank Shandong University of Science and Tech-
nology for its generous support of ICIMCS 2017. Perched on the southern tip of the
Shandong Peninsula, Qingdao is well-known to the West by its postal map spelling of
Tsingtao. It is known as “Pearl on the Yellow River” and “China’s Switzerland.” The
most famous beer of China—Tsingtao Beer—is produced here. We hope that our
venue made ICIMCS 2017 a memorable experience for all participants.
We received 103 paper submissions, covering topics of multimedia information
fusion, image processing and object recognition, machine learning and representation
learning, as well as multimedia retrieval. The submitted papers were reviewed by the
Technical Program Committee, consisting of 68 reviewers. Each paper was reviewed
by at least two reviewers. The program chairs carefully considered the input and
feedback from the reviewers and accepted 48 papers (46.6%) to the main conference.
Among the accepted papers, 20 were oral papers and 28 were poster papers. This
volume of the conference proceedings contains all the oral and poster papers.
The technical program is an important aspect but only achieves its full impact if
complemented by challenging keynotes. We are extremely pleased and grateful to have
had three exceptional keynote speakers, Bin Hu, Jingdong Wang, and Xiansheng Hua,
accept our invitation and present interesting ideas and insights at ICIMCS 2017.
We are heavily indebted to many individuals for their significant contributions.
First, we are very grateful to all the authors who contributed their high-quality research
and shared their knowledge with our scientific community. Second, we wish to thank
all Organizing and Program Committee members, reviewers, session chairs, student
volunteers, and supporters. Their contributions are much appreciated. We hope you all
enjoy the proceedings of ICIMCS 2017.

September 2017 Benoit Huet

Liqiang Nie
Richang Hong
 Organization

General Chairs
Alex Hauptmann Carnegie Mellon University, USA
Baoquan Chen Shandong University, China
Donghua Zhou Shandong University of Science and Technology, China

Technical Program Chairs

Benoit Huet EURECOM, France
Liqiang Nie Shandong University, China
Richang Hong Hefei University of Technology, China

Sponsorship Chair
Xinshun Xu Shandong University, China

Publication Chairs
Tian Gan Shandong University, China
Hanwang Zhang Columbia University, USA

Special Session Chairs

Liang Li Chinese Academy of Sciences, China
Xuemeng Song Shandong University, China

Registration Chair
Weifeng Liu China University of Petroleum (East China), China

Finance Chair
Weijian Ni Shandong University of Science and Technology, China

Local Chair
Qingtian Zeng Shandong University of Science and Technology, China

Publicity Chair
Yongquan Liang Shandong University of Science and Technology, China
VIII Organization

Program Committee
Lamberto Ballan University of Florence, Italy
Kseniya Buraya National University of Singapore, Singapore
Da Cao Xiamen University, China
Jingjing Chen City University of Hong Kong, Hong Kong, SAR China
Jingyuan Chen National University of Singapore, Singapore
Zhumin Chen Shandong University, China
Wen-Huang Cheng Academia Sinica, Taiwan
Zhiyong Cheng National University of Singapore, Singapore
Wei-Ta Chu National Chung Cheng University, Taiwan
Noel Codella IBM Research, USA
Chaoran Cui Shandong University of Finance and Economics, China
Yuming Fang Nanyang Technological University, Singapore
Aleksandr Farseev National University of Singapore, Singapore
Fuli Feng National University of Singapore, Singapore
Tian Gan Shandong University, China
Ke Gao Institute of Computing Technology, Chinese, China
Lianli Gao The University of Electronic Science and Technology
of China, China
Zan Gao Tianjing University of Technology, China
Francesco Gelli National University of Singapore, Singapore
Xue Geng National University of Singapore, Singapore
Kailing Guo University of Technology Sydney, Australia
Jungong Han Civolution Technology, The Netherlands
Shijie Hao HeFei University of Technology, China
Steven Hoi Singapore Management University, Singapore
Wolfgang Huerst Utrecht University, The Netherlands
Muwei Jian Ocean University of China, China
Tae Hoon National University of Singapore, Singapore
Joseph Kim
Duy-Dinh Le National Institute of Informatics, Japan
Liang Li University of Chinese Academy of Sciences, China
Xirong Li Renmin University of China, China
Zechao Li Nanjing University of Science and Technology, China
Zhen Li University of Illinois Urbana-Champaign, USA
Anan Liu Tianjing University, China
Bo Liu Rutgers University, USA
Meng Liu Shandong University, China
Weifeng Liu China University of Petroleum (East China), China
Xueliang Liu Hefei University of Technology, China
Zhu Liu AT&T Labs Research, USA
He Ma Northeastern University, China
Qirong Mao Jiangsu University, China
Yadong Mu Columbia University, USA
Chong-Wah Ngo City University of Hong Kong, Hong Kong, SAR China
 Organization IX

Weijian Ni Shandong University of Science and Technology, China

Weizhi Nie Tianjing University, China
Xiushan Nie Shandong University of Finance and Economics, China
Tongwei Ren Nanjing University, China
Jitao Sang Chinese Academy of Sciences, China
Xiangjun Shen Jiangsu University, China
Xiangbo Shu Nanjing University of Science and Technology, China
Jiande Sun Shandong University, China
Sheng Tang University of Chinese Academy of Sciences, China
Qi Tian University of Texas at San Antonio, USA
Xiang Wang National University of Singapore, Singapore
Xiangyu Wang Facebook, USA
Xiaochi Wei Beijing Institute of Technology, China
Lynn Wilcox FXPAL, USA
Yang Wu Kyoto University, Japan
Hongtao Xie University of Chinese Academy of Sciences, China
Xinshun Xu Shandong University, China
Feng Xue HeFei University of Technology, China
Yi Yang University of Technology Sydney, Australia
Chunyun Zhang Shandong University of Finance and Economics, China
Jianglong Zhang Communication University of China, China
Wanlei Zhao Inria, France
Lei Zhu The University of Queensland, Australia
 Contents

Best Paper

Joint Deep Learning and Gaussian Representation

for Person Re-identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Nan Song, Xianglei Zhu, and Yahong Han

Blind Image Deblurring Using Adaptive Priors . . . . . . . . . . . . . . . . . . . . . . 13

Bingwang Zhang, Risheng Liu, Haojie Li, Qi Yuan, Xin Fan,
and Zhongxuan Luo

Joint Latent Space and Multi-view Feature Learning . . . . . . . . . . . . . . . . . . 23

Kailing Guo, Xiangmin Xu, Bolun Cai, and Tong Zhang

Sketch-Based Image Retrieval with Multiple Binary HoG Descriptor . . . . . . . 32

Tianqi Wang, Liyan Zhang, and Jinhui Tang

Multimedia Information Fusion

Joint Visual Context for Pedestrian Captioning . . . . . . . . . . . . . . . . . . . . . . 45

Quan Liu, Yingying Chen, Jinqiao Wang, and Sijiong Zhang

Choose the Largest Contributor: A Fusion Coefficient Learning Network

for Semantic Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Jianzhuang Yu, Shuai Zhao, and Yahong Han

Clothing Matching Based on Multi-modal Data . . . . . . . . . . . . . . . . . . . . . 65

Zekun Li, Xuemeng Song, Tian Gan, Zhumin Chen, and Xiaoyu Zhu

Bidirectional Multimodal Recurrent Neural Networks with Refined Visual

Features for Image Captioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Yanwu Shu, Liyan Zhang, Zechao Li, and Jinhui Tang

Image Processing and Object Recognition

Automatic Watermeter Digit Recognition on Mobile Devices . . . . . . . . . . . . 87

Yunze Gao, Chaoyang Zhao, Jinqiao Wang, and Hanqing Lu

3D Human Body Reshaping with Anthropometric Modeling. . . . . . . . . . . . . 96

Yanhong Zeng, Jianlong Fu, and Hongyang Chao
XII Contents

Learning Multi-Scale Shrinkage Fields for Blind Image Deblurring . . . . . . . . 108

Bingwang Zhang, Risheng Liu, Haojie Li, Qi Yuan, Xin Fan,
and Zhongxuan Luo

Indoor RGB-D Object Detection with the Guidance of Hand-Held Objects. . . 116
Leixian Qiao, Yaohui Zhu, Runze Li, Weiqing Min, and Shuqiang Jiang

Machine Learning and Representation Learning

Building Weighted Classifier Ensembles Through Classifiers Pruning . . . . . . 131

ChenWei Cai, Dickson Keddy Wornyo, Liangjun Wang,
and XiangJun Shen

Elastic Net Regularized Dictionary Learning for Face Recognition . . . . . . . . 140

Li Wang, Yan-Jiang Wang, and Bao-Di Liu

Extracting Deep Video Feature for Mobile Video Classification

with ELU-3DCNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Jihong Liu, Jing Zhang, Hui Zhang, Xi Liang, and Li Zhuo

A Robust End-to-End Neural Network for Scene Text Detection. . . . . . . . . . 160

Qi Yuan, Haojie Li, Zhihui Wang, Xin Fan, and Zhongxuan Luo

Multimedia Retrieval

Exploiting Concept Correlation with Attributes for Semantic Binary

Representation Learning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Haiping Wu, Yang Yang, Xing Xu, Fumin Shen, Ning Xie, and Yanli Ji

Image Retrieval Based on Optimized Visual Dictionary

and Adaptive Soft Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Hui Liu, Zechao Li, and Xiangbo Shu

Query Modeling for Click Data Based Image Recognition Using Graph
Based Propagation and Sparse Coding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
Weichen Wu, Min Tan, Guangjian Zheng, and Jun Yu

Zero-Shot Cross-Media Retrieval with External Knowledge . . . . . . . . . . . . . 200

Jingze Chi, Xin Huang, and Yuxin Peng

Regular Poster Papers

A Novel Computer-Aided Diagnosis Method of Nasopharyngeal

Carcinoma Based on Magnetic Resonance Images. . . . . . . . . . . . . . . . . . . . 215
Xiang Tian, Yan Zhang, Qingbin Wu, Changzheng Shi, Xiaoping Li,
Chunmei Qing, and Lin Shu
 Contents XIII

A Real-Time Distributed Index Based on Topic

for Microblogging System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
Zhikun Chen, Lu Wang, and Shuqiang Yang

Auto-Segmentation of Pathological Lung Parenchyma Based on Region

Growing Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Jiyang Dong, Ke Lu, Shuangfeng Dai, Jian Xue, and Rui Zhai

Classification of Heart Sounds Based on Convolutional Neural Network . . . . 252

Tianya Li, Chunmei Qing, and Xiang Tian

Content-Based Similar Document Image Retrieval Using Fusion

of CNN Features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Mao Tan, Siping Yuan, and Yongxin Su

Continuous Dropout Strategy for Deep Learning Network . . . . . . . . . . . . . . 271

Jianchao Fei, Ting Rui, Xiaona Song, You Zhou, and Sai Zhang

Cross-Media Semantic Matching via Sparse Neural Network

Pre-trained by Deep Restricted Boltzmann Machines . . . . . . . . . . . . . . . . . . 280
Bin Zhang, Huaxiang Zhang, Jiande Sun, Zhenhua Wang,
Hongchen Wu, and Xiao Dong

Deep Auto-Encoder Based on Supervised Learning for Damaged

Face Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
Ting Rui, Sai Zhang, Junhua Zou, You Zhou, and Jian Tang

Design and Implementation of Secure Multimedia Storage System . . . . . . . . 300

Guang-yu Gu, Qing Yi, Hao Zheng, Shu-juan Zhang, and Bo-wu Sun

Emotion Recognition from EEG Using RASM and LSTM . . . . . . . . . . . . . . 310

Zhenqi Li, Xiang Tian, Lin Shu, Xiangmin Xu, and Bin Hu

Exemplar-Based Photo Color Correction by Exploring Visual Aesthetics . . . . 319

Zhenkun Zhou, Shijie Hao, and Mingyu Liu

Fast Nearest Neighbor Search Based on Approximate k-NN Graph . . . . . . . . 327

Jie Yang, Wan-Lei Zhao, Cheng-Hao Deng, Hanzi Wang,
and Sangwhan Moon

Image Recommendation on Content-Based Bipartite Graph . . . . . . . . . . . . . 339

Lifang Wu, Lei Zhang, Meng Jian, Dai Zhang, and Haiying Liu

Initialized Frame Attention Networks for Video Question Answering. . . . . . . 349

Kun Gao, Xianglei Zhu, and Yahong Han

MFCC: An Efficient and Effective Matrix Factorization Model

Based on Co-clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
Wenjuan Yang, Le Wu, Xueliang Liu, and Chunxiao Fan
XIV Contents

Multi-layers CNNs for 3D Model Retrieval . . . . . . . . . . . . . . . . . . . . . . . . 371

Anan Liu, Shu Xiang, Weizhi Nie, and Yuting Su

Multispectral Image Denoising Based on Non-local Means

and Bilateral Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Xueyan Zhen, Ning He, Xin Sun, and Yuqing Zhang

Probabilistic Collaborative Representation with Kernels

for Visual Classification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
Jie Meng, Yanjiang Wang, and Bao-Di Liu

Probability Matrix SVM+ Learning for Complex Action Recognition . . . . . . 403

Fang Liu, Xiangmin Xu, Chunmei Qing, and Jianxiu Jin

Progressive Lifelong Learning by Sharing Representations

for Few Labeled Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Guoxi Su, Xiangmin Xu, Chaowen Chen, Bolun Cai,
and Chunmei Qing

Refining Eye Synthetic Images via Coarse-to-Fine Adversarial Networks

for Appearance-Based Gaze Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
Tongtong Zhao, Yafei Wang, and Xianping Fu

Robust Steganography via Patch-Based Texture Synthesis . . . . . . . . . . . . . . 429

Zhengcai Qin, Meng Li, and Bin Wu

Saliency Cuts on RGB-D Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440

Yuantian Wang, Lei Huang, Tongwei Ren, and Yunfei Zhang

Same-Style Products Mining for Clothes Retrieval . . . . . . . . . . . . . . . . . . . 449

Zhenwei Shen, Zhiwei Fang, and Jing Liu

Scattering Wavelet Based Deep Network for Image Classification . . . . . . . . . 459

Wenyou Huang, Xiaoyi Zou, and Rui Qiao

Scene Text Detection with Cascaded Filtering and Grouping Modules . . . . . . 470
Lifei Zhang and Xinguang Xiang

Towards Underwater Image Enhancement Using Super-Resolution

Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479
Xueyan Ding, Yafei Wang, Zheng Liang, Jun Zhang, and Xianping Fu

Traffic Sign Recognition with Inception Convolutional Neural Networks . . . . 487

Jianyun Lu, Yanyun Qu, and Xiaodong Yang

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495

Best Paper
 Joint Deep Learning and Gaussian
Representation for Person
Re-identification

Nan Song1(B) , Xianglei Zhu2 , and Yahong Han1

1
School of Computer Science and Technology, Tianjin University, Tianjin, China
{songnan,yahong}@tju.edu.cn
2
China Automotive Technology and Research Center, Tianjin, China
zhuxianglei@catarc.ac.cn

Abstract. Person re-identification (re-id) has attracted extensive atten-

tion worldwide in the computer vision. Unlike other images retrieval
tasks, the color and texture information of pedestrian clothing is an
important clue of person re-id. However, these special features of re-id
are not focused in the process of extracting features by deep learning. We
find that some traditional methods of extracting features are better than
deep learning in color and texture. Therefore, in this paper, we propose
a model of jointing Deep learning and Gaussian representation (JDAG),
which is the fusion of CNN network and Gaussian descriptor to enrich
the person re-id features. We also adjust the parameters of our model to
achieve an optimal performance and obtain more discriminative features.
The experimental results generated on two representative and public
datasets (a large-scale dataset Market1501 and a small dataset VIPeR)
are evaluated using three different distance metric learning strategies.
The comparisons show that our method is superior than other existing
methods.

Keywords: Deep learning · Fusion strategy · Person re-identification

1 Introduction

Recent years, person re-identification (re-id) is a new technology in computer

vision and intelligent surveillance systems, which is mainly applied on person
recognition and pedestrian similarity matching in a complex and changeful envi-
ronment. The key of problems is described as a task of matching same person
ID between gallery images and index query images from different camera views.
However, intra-personal variations such as blur, illumination, occlusion, per-
son pose, resolution, and the images intercepted from different camera views,
bring great challenges. Therefore, researchers are concerned with improving the
accuracy and efficiency in recognition and retrieval. Some effective methods are
proposed [1–10].

c Springer Nature Singapore Pte Ltd. 2018

B. Huet et al. (Eds.): ICIMCS 2017, CCIS 819, pp. 3–12, 2018.
https://doi.org/10.1007/978-981-10-8530-7_1
4 N. Song et al.

Feature extraction, similarity computation and evaluation are three main

steps to solve the person re-id problems generally. Firstly, the convolutional
neural network (CNN) has been widely applied on image feature extraction and
achieved state-of-the-art results [11–13]. However, there are two difficulties: (1)
Training set of re-id is relatively small so that it is prone to over-fitting. (2) Re-id
problem is mainly based on the color and texture information of person. While
these characteristics are not mainly focused by CNN, instead, semantic features
are extracted. Although semantic features may be more accurate and rich, small
datasets have a better performance for special features using traditional machine
learning, which motivates us to combine them to improve features discrimination.
As known, the features from traditional machine learning are also studied as a
hot spot at present, included Local Maximal Occurrence (LOMO) features [1],
Covariance-of-Covariance feature [14], Gaussian Of Gaussian (GOG) features
recently proposed in [4] which can combine mean and covariance information in
region hierarchies and achieve the state-of-the-art performances, so we regard it
as traditional branches in our work.
Similarity computation and evaluation can be treat as a whole by distance
metrics learning (DML) for a pair of images. At present, there are many DML
methods which have been proved to be effective for person re-id. Davis et al.
[15] presented an information-theoretic approach (ITML) to handle a variety
constraints problem and Guillaumin et al. [16] raised a logistic discriminant
method to learn the labeled image pairs (LDM). In addition of that, it also
includes: person re-id methods (eSDC) [3], KISSME [17], XQDA [1] and using
Euclidean distance to compute the score of similarity directly. In our paper, we
used the last three methods to evaluate results and the contrast results also were
listed in experiment section.
As mentioned above, a small dataset trained by deep learning model such as
ResNet [18] gets a bad result due to over-fitting. Therefore, traditional meth-
ods had been employed in the high-dimensional feature vectors to express the
person features of small datasets. However, these features need to be extracted
based on special attributes (color, texture), so that it is more deficient to extract
discriminative features than deep learning. In order to solve the problems, we
consider the color and texture of pedestrian clothing as a focus jointing the
CNN features and training in ImageNet model. The fusion can not only solve
the problem of over-fitting, but also solve the attention problem of color and
texture information.
In this paper, the important contributions of our work are: (1) we discover the
issue that small datasets (VIPeR [19]) tend to over-fitting in CNN model. How-
ever, the special features of color and texture enhanced by traditional machine
learning have applicability to solve this problem of person re-id. (2) We propose a
fusion model of jointing Deep learning and Gaussian representation (JDAG). The
traditional features and CNN features are trained in the network by fusion strat-
egy. (3) Our method is tested on two benchmarks (a large dataset Market1501
[20] and a small dataset VIPeR) and achieved a good performance. Besides, we
use the three kinds of relatively common DML strategies (XQDA, KISSME,
Euclidean distance) to find out a most suitable one for our JDAG model.
 Joint Deep Learning and Gaussian Representation for Person Re-id 5

The organization of this paper is as follows: Firstly, we describe the two

branches of JDAG model about the two features in fusion respectively and how
to concatenate the features in a 2stream siamese network [21] in Sect. 2. In
Sect. 3, we present our experimental details and show the comparison results on
representative datasets. At last, we conclude our method in Sect. 4.

2 Methodology
In this section, we introduce our network architecture. At first, we extract the
GOG features of four different color channels (HSV, Lab, RnG, RGB) for a pair
of images as the traditional features inputs, then the mean and L2 normalization
can be used next. Secondly, the features which are fusions of GOG and CNN for
a pair of images are fed into the identification network and verification network
to train together simultaneously. Finally, we test and evaluate our method in
public datasets using state-of-the-art metric learning strategies.

2.1 Branch of CNN Features

In the previous works of person re-id, verification models and identification mod-
els are trained separately. Verification models are mainly used to calculate the
similarity of images, which allow a pair of images as inputs and then similarity
scores are obtained by cosine distance between the features of outputs in order
to predict whether the IDs are same or not. Identification models are developed
for classification and the number of categories varies with the number of person
IDs in datasets, so it can be treated as a large-scale multi-classification problem
where an image as input can be predicted for the corresponding person ID.
It is difficult to improve the accuracy of similarity scores, because of only con-
sidering optimizing verification models and that person ID can be only treated as
a ground truth label of loss function in the previous works. Recently, a method
of combining the two models was proposed [21] to take advantages of both net-
works and form a unified framework, so we select this network to train datasets.
However, the difference is that ours method adds GOG features as input data
to predict person ID and calculate similarity in stead of only using the out-
puts from the pool5 layer in ResNet50. It is necessary to remove mean and L2
normalization before GOG features are fed into the network.
The architecture of JDAG model is shown in Fig. 1: we introduce the 2stream
siamese of Resnet50 to extract the features of pool5 layer for the inputs of pairs.
Meanwhile, the traditional features represented by GOG descriptor fg and fg 2
can be fed into the respective FC layers and the outputs are indicated by f1
and f2 respectively. Next, we combine the pool5 features and GOG features into
fusion features indicated by f1∗ and f2∗ . On the one hand, the fusion features are
applied on identifying person ID for the pairs of images by FC layers of multi
classification. On the other hand, we use the Euclidean distance layer to compare
the fusion features and predict similarity scores by the next FC layer of binary
classification. The Euclidean distance layer is denoted as
f = (f1∗ − f2∗ )2 (1)
6 N. Song et al.

Fig. 1. JDAG model structure. Positive or Negative pairs of images of size 256 × 256
input ResNet50 network and GOG model and the features of pool5 and fg are outputs
respectively. The fusion features which achieved by FC layers and Concat layers can
be fed into Identification model and Verification model. Finally, the three loss can be
obtained.

Finally, two identification losses and a verification loss computed by softmax

layers are used back propagation. Identification loss and verification loss can be
computed by cross-entropy loss which is
n

H(p, q) = − p(x)logq(x) (2)
i=1

x is the features vector. p(x) and q(x) are the probability of predicting and target
by softmax layer respectively and H is the cross-entropy loss.

2.2 Branch of GOG Features

We select GOG feature as traditional features to combine with CNN features,

because GOG feature is more concerned about the color and texture attributes
of pedestrian images, which is more suitable for person features extraction. We
briefly review the GOG features extraction method as follow: firstly, an image
is divided into n regions. M patches of size k × k are selected in a region. Then,
a 8 dimensional feature vector fi is extracted for each pixel i in a patch.

fi = [y, M0◦ , M90◦ , M180◦ , M270◦ , R, G, B]T

y denotes the features of the vertical direction. Due to that people walk upright
normally on the road, the person images tend to be consistent in the vertical
direction but great changeful in the horizontal when pose or view is changed. M
denotes gradient magnitude for different angles which can describe the texture
information of person clothing effectively. R, G, B denote the color information
of clothing, so they are the most important cue for person re-id. Therefore, four
different color channels are selected to extend the color space, which include
 Joint Deep Learning and Gaussian Representation for Person Re-id 7

HSV, Lab, RnG and RGB. After extracting the features of the pixels, each pixel
information is summarized in a patch feature vector by Gaussian distribution.
Next, a region feature can be represented by summarizing all the patches features
with Gaussian expression. At last, n regions features are concatenated to describe
the whole image. So, the feature is called Gaussian of Gaussian (GOG) feature.
It is not difficult to find that GOG feature is mainly based on color and tex-
ture attributes, and this is a state-of-the-art method for person feature extrac-
tion. After extracting the GOG feature, removing mean value and L2 normaliza-
tion are utilized at next stage in order to make it easier to combine with CNN
features. For more algorithmic details, please refer to [4].

2.3 Strategies of Fusion

Processing Data with PCA. An image processed by the GOG method can
transform into a 1 × 7567 dimensional feature vector and in the CNN branch, can
transform into a 1 × 1 × 2048 dimensional features in the pool5 layer of resnet50
network. In order to balance the GOG features and CNN features and reduce
impacts for redundancy results, PCA method for reducing dimension is applied
on our work. Before the GOG features are fed into the ResNet50, the dimension
can be handled by PCA, so that the dimension is 1 × 1 × DP CA .
The advantages of processing GOG features by PCA are reduction of features
dimension and effective balance between GOG and CNN. What’s more, PCA can
reduce parameters for the network and improve the speed for training. However,
the disadvantage is that magnifying the dimensions of noise which causes a small
standard deviation so that the deviation of whole is increased.

Processing Data with Full Connection. As above section, although PCA

method is helpful for reducing dimension of features, the deviations can be mag-
nified. So, we propose to replace the PCA with a fully connected layer to reduce
the dimension. In fact, the fully connection layer can be regarded as a spe-
cial convolution layer. After the GOG features are processed by reshape layer,
1 × 1 × 7567 × DF C filter of FC layer can be set and the concatenation of outputs
and pool5 features is fed into identification layer and verification layer as their
inputs. In our experiment, the DF C filter size is changed to balance the fusion
feature.
The advantages of FC layer are that all the information of the GOG features
is retained, and the weights are learned well by training the network. But the
speed of training is slow because of the parameters increased.

3 Experimental Results
In order to verify the effectiveness of our proposed method, we make experiments
on a large scale dataset market1501 and a small dataset VIPeR. We report
the results which are compared with the current state-of-the-art results on two
datasets and show that our method performs best.
8 N. Song et al.

Table 1. Impact of different dropout parameters on Market1501 dataset in three DML.

D Our + Euclidean Our + XQDA Our + KISSME

MAP Rank-1 MAP Rank-1 MAP Rank-1
0.4 64.7 83.4 64.2 82.1 65.6 84.8
0.5 65.7 84.4 65.0 83.2 66.0 84.8
0.6 65.7 83.8 64.8 81.7 65.9 83.9
0.7 66.1 83.9 65.2 82.2 66.0 84.2
0.8 65.4 83.8 64.8 81.7 65.9 84.0
0.9 61.9 81.3 62.7 81.2 63.8 82.9

Market1501 dataset contains 32,668 detected bounding boxes of 1,501 per-

son IDs and the images corresponded each ID are captured by DPM method
from the videos of 6 cameras. The dataset is divided into three parts: the train-
ing set consists of 12,936 cropped images and 751 person IDs, the test set consists
of 19,732 cropped images and 750 person IDs, and a single query set includes
3,368 person images. The task is to select the ground truth images of the corre-
sponding query set from all candidate images of the test set. The settings of the
training and testing process are same as [22] in single query and multiple query
evaluated.
We train the 2stream siamese JDAG model which is consisted of two
ResNet50 and fine-tune it on ImageNet. FC layer filter size is set to 1 × 1 ×
7567 × 256. The dropout parameters are adjusted as shown in Table 1 to reduce
the the occurrence of over-fitting phenomenon. We report the mean average
precision (MAP) and Rank-1 on Market1501 in three different DML when the
dropout parameter is changed from 0.4 to 0.9. It is not difficult to find that
when dropout is set to 0.7, our results perform best by the Euclidean distance
and KISSME.
We also compare our method with other state-of-the-art method which are
published within a year in terms of MAP and Rank-1 on Market1501 dataset
in single query and multi query as Table 2. The dropout fixed is set to 0.7.
(a) group denotes the latest results obtained by the existing method in a year.
(b) and (c) groups denote the results evaluating the pool5 layer features and
fusion layer features, respectively. We note that compared to the result obtained
without using the person special features, our result has a 6% advantage on
hard level. (d) group denotes the results adding re-ranking method [23], we were
surprised to find that MAP of our results have improved about 13% and the rank-
1 has a 3% advantage than the results only using CNN features and re-ranking
in the Euclidean distance and the XQDA evaluation strategies. According the
comparison of Table 2, our method performs better than the state-of-the-art
results no matter which evaluation strategies.
VIPeR is a challenging dataset which includes 1,264 person images and 632
person IDs from two cameras. Because the scale of training set is small, the
 Joint Deep Learning and Gaussian Representation for Person Re-id 9

Table 2. Comparison with state-of-the-art results of person re-identification on Mar-

ket1501 dataset. (a) Previous works and results. (b) Our results by evaluating the
features of pool5 layer (c) Our results by evaluating the features of fusion. The MAP
and Rank-1 are listed for single query and multi query, respectively.

Method Year of Publication Single Query Multi Query

MAP Rank-1 MAP Rank-1
(a) Gate Reid ECCV 2016 [24] 39.6 65.9 48.5 76.0
CRAFT − MFA + LOMO TPAMI 2017 [25] 45.5 71.8 54.3 79.7
SOMAnet Arxiv 2017 [26] 47.9 73.9 57.0 81.3
IDE + Re-ranking CVPR 2017 [23] 63.6 77.1 - -
PIE descriptor + KISSME Arxiv 2017 [27] 56.0 79.3 - -
Re-id based CNN Embedding Arxiv 2016 [21] 59.9 79.5 70.3 85.8
SVDNet Arxiv 2017 [28] 62.1 82.3 - -
Deep transfer learning Arxiv 2016 [29] 65.5 83.7 73.8 89.6
(b) JDAG(pool5) + Euclidean Ours 66.5 84.4 72.3 86.1
JDAG (pool5) + XQDA Ours 65.0 82.1 70.2 85.1
JDAG (pool5) + KISSME Ours 65.9 83.4 71.0 85.4
(c) JDAG (fusion) + Euclidean Ours 66.1 83.4 71.5 85.9
JDAG (fusion) + XQDA Ours 65.2 82.2 71.0 84.9
JDAG (fusion) + KISSME Ours 66.0 84.2 71.6 85.7
(d) JDAG (fusion) + Euclidean + Re-ranking Ours 79.3 86.3 81.2 88.5
JDAG (fusion) + XQDA + Re-ranking Ours 76.7 84.0 79.4 87.0

Table 3. Comparison of state-of-the-art results on VIPeR dataset.

Method Rank-1 Rank-5 Rank-10 Rank-20

MLFL [11] 29.1 - 65.9 70.9
SCNCD [30] 37.8 68.5 81.2 90.4
Semantic [12] 31.1 68.6 82.8 94.9
LOMO + XQDA [1] 40.0 - 80.5 91.1
MetricEnsemble [13] 45.9 77.5 88.9 95.8
GOG + XQDA [4] 49.7 79.7 88.7 94.5
Ours + XQDA 57.2 84.5 91.5 96.2

features are extracted by Market1501 training model. Then, we randomly select

316 images for testing and evaluation by XQDA and the procedure is conducted
for 10 splits. The comparison with other method is reported in Table 3. We can
observe that our result is superior to existing results. The precision is improved
by 8% in Rank-1 than only using GOG features.
In order to explain this result intuitively, we choose an example in
VIPeR dataset to compare the accuracy of the matching between GOG +
XQDA and JDAG + XQDA, as Fig. 2. A ground truth image from a query is
ranked the fourth position in GOG + XQDA, however it is ranked the first in
JDAG + XQDA, which fully shows that our performance is more outstanding.
10 N. Song et al.

Fig. 2. Person re-id samples on VIPeR dataset. The images in left column is query
images. The tested images are ranked according to the similarity scores. The retrieval
images in top row are from GOG method and the bottom is from our method. The red
boxes denote error detection and the green boxes denote exactness. (Color figure online)

The result above on two datasets indicate that discriminative features

extracted by JDAG model are more suitable for application to the large scale
datasets and small datasets.

4 Conclusions
In this work, we introduce a novel model called JDAG which is a fusion network
of 2stream ResNet50 and GOG. When the JDAG model is trained, the tradi-
tional GOG features is fed into the model to play an important role in empha-
sizing special features. The features trained and extracted by JDAG model are
more discriminative than the single features extracted. We utilize the JDAG
features to improve the robustness of matching score. We verify the results on
two public datasets and achieve a state-of-the-art performance. In the future
work, we will continue to explore the robust feature extraction method based
on JDAG model. We also consider applying our method on more areas, such as
person search, re-id in video and some related retrieval tasks.

Acknowledgment. This work was supported by the NSFC (under Grant U1509 206,
61472276).

References
1. Liao, S., Hu, Y., Zhu, X., Li, S.Z.: Person re-identification by local maximal occur-
rence representation and metric learning. In: Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, pp. 2197–2206 (2015)
2. Zhao, R., Ouyang, W., Wang, X.: Person re-identification by salience matching.
In: Proceedings of the IEEE International Conference on Computer Vision, pp.
2528–2535 (2013)
3. Zhao, R., Ouyang, W., Wang, X.: Unsupervised salience learning for person re-
identification. In: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 3586–3593 (2013)
 Joint Deep Learning and Gaussian Representation for Person Re-id 11

4. Matsukawa, T., Okabe, T., Suzuki, E., Sato, Y.: Hierarchical Gaussian descriptor
for person re-identification. In: Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 1363–1372 (2016)
5. Hong, R., Zhang, L., Zhang, C., Zimmermann, R.: Flickr circles: aesthetic ten-
dency discovery by multi-view regularized topic modeling. IEEE Trans. Multimedia
18(8), 1555–1567 (2016)
6. Hong, R., Hu, Z., Wang, R., Wang, M., Tao, D.: Multi-view object retrieval via
multi-scale topic models. IEEE Trans. Image Process. 25(12), 5814–5827 (2016)
7. Hong, R., Yang, Y., Wang, M., Hua, X.S.: Learning visual semantic relationships
for efficient visual retrieval. IEEE Trans. Big Data 1(4), 152–161 (2015)
8. Nie, L., Yan, S., Wang, M., Hong, R., Chua, T.S.: Harvesting visual concepts for
image search with complex queries. In: Proceedings of the 20th ACM international
conference on Multimedia, pp. 59–68 (2012)
9. Nie, L., Wang, M., Zha, Z., Li, G., Chua, T.S.: Multimedia answering: enriching
text QA with media information. In: Proceedings of the 34th International ACM
SIGIR Conference on Research and Development in Information Retrieval, pp.
695–704 (2011)
10. Nie, L., Wang, M., Zha, Z.J., Chua, T.S.: Oracle in image search: a content-based
approach to performance prediction. ACM Trans. Inf. Syst. (TOIS) 30(2), 13
(2012)
11. Zhao, R., Ouyang, W., Wang, X.: Learning mid-level filters for person re-
identification. In: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 144–151 (2014)
12. Shi, Z., Hospedales, T.M., Xiang, T.: Transferring a semantic representation for
person re-identification and search. In: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 4184–4193 (2015)
13. Paisitkriangkrai, S., Shen, C., van den Hengel, A.: Learning to rank in person re-
identification with metric ensembles. In: Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 1846–1855 (2015)
14. Serra, G., Grana, C., Manfredi, M., Cucchiara, R.: Covariance of covariance fea-
tures for image classification. In: Proceedings of International Conference on Mul-
timedia Retrieval, p. 411 (2014)
15. Davis, J.V., Kulis, B., Jain, P., Sra, S., Dhillon, I.S.: Information-theoretic metric
learning. In: Proceedings of the 24th International Conference on Machine Learn-
ing, pp. 209–216 (2007)
16. Guillaumin, M., Verbeek, J., Schmid, C.: Is that you? metric learning approaches
for face identification. In: 2009 IEEE 12th International Conference on Computer
Vision, pp. 498–505 (2009)
17. Koestinger, M., Hirzer, M., Wohlhart, P., Roth, P.M., Bischof, H.: Large scale met-
ric learning from equivalence constraints. In: 2012 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pp. 2288–2295 (2012)
18. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 770–778 (2016)
19. Gray, D., Tao, H.: Viewpoint invariant pedestrian recognition with an ensemble
of localized features. In: Forsyth, D., Torr, P., Zisserman, A. (eds.) ECCV 2008.
LNCS, vol. 5302, pp. 262–275. Springer, Heidelberg (2008). https://doi.org/10.
1007/978-3-540-88682-2 21
20. Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., Tian, Q.: Scalable person re-
identification: a benchmark. In: Proceedings of the IEEE International Conference
on Computer Vision, pp. 1116–1124 (2015)
12 N. Song et al.

21. Zheng, Z., Zheng, L., Yang, Y.: A discriminatively learned CNN embedding for
person re-identification. arXiv preprint arXiv:1611.05666 (2016)
22. Zagoruyko, S., Komodakis, N.: Learning to compare image patches via convo-
lutional neural networks. In: Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 4353–4361 (2015)
23. Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking Person Re-identification with
k-reciprocal Encoding. arXiv preprint arXiv:1701.08398 (2017)
24. Varior, R.R., Haloi, M., Wang, G.: Gated siamese convolutional neural network
architecture for human re-identification. In: Leibe, B., Matas, J., Sebe, N., Welling,
M. (eds.) ECCV 2016. LNCS, vol. 9912, pp. 791–808. Springer, Cham (2016).
https://doi.org/10.1007/978-3-319-46484-8 48
25. Chen, Y.C., Zhu, X., Zheng, W.S., Lai, J.H.: Person re-identification by camera
correlation aware feature augmentation. IEEE Trans. Pattern Anal. Mach. Intell.
(2017)
26. Barbosa, I.B., Cristani, M., Caputo, B., Rognhaugen, A., Theoharis, T.: Looking
beyond appearances: synthetic training data for deep CNNs in re-identification.
arXiv preprint arXiv:1701.03153 (2017)
27. Zheng, L., Huang, Y., Lu, H., Yang, Y.: Pose invariant embedding for deep person
re-identification. arXiv preprint arXiv:1701.07732 (2017)
28. Sun, Y., Zheng, L., Deng, W., Wang, S.: SVDNet for Pedestrian Retrieval. arXiv
preprint arXiv:1703.05693 (2017)
29. Geng, M., Wang, Y., Xiang, T., Tian, Y.: Deep transfer learning for person re-
identification. arXiv preprint arXiv:1611.05244 (2016)
30. Yang, Y., Yang, J., Yan, J., Liao, S., Yi, D., Li, S.Z.: Salient color names for person
re-identification. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV
2014. LNCS, vol. 8689, pp. 536–551. Springer, Cham (2014). https://doi.org/10.
1007/978-3-319-10590-1 35
 Blind Image Deblurring Using
Adaptive Priors

Bingwang Zhang1,2 , Risheng Liu1,2 , Haojie Li1,2(B) , Qi Yuan1,2 , Xin Fan1,2 ,

and Zhongxuan Luo1,2
1
DUT-RU International School of Information and Software Engineering,
Dalian University of Technology, Dalian, China
bwzhang@mail.dlut.edu.cn , hjli@dlut.edu.cn
2
Key Laboratory for Ubiquitous Network
and Service Software of Liaoning Province, Dalian, China

Abstract. For blind image deblurring, a good prior knowledge can guide
the maximum a posterior (MAP) based algorithms to be away from the
trivial solution. Therefore, many existing methods focus on designing
effective priors to constrain the solution space. However, blind deconvo-
lution with fixed priors is not robust. And many priors are extremely
costly to design and compute. In this paper, we proposed a blind decon-
volution method with adaptive priors under the MAP framework. Specif-
ically, we carry out our algorithm under the multi-scale, and at each scale
we add specific sparse regularization to standard deblurring formulation.
By tunning both the priors and the weights we can give more flexible
sparse regularization constraint. After iteration, our algorithm output
both latent image and estimated blur kernel, simultaneously. We prove
the convergence of the proposed algorithm. Extensive experiments show
the effectiveness of our proposed approach.

Keywords: Blind deblurring · Sparsely constraint · Adaptive priors

MAP

1 Introduction
Blind image deblurring aims to recover a latent sharp image and a blur ker-
nel from a blurred image. And it has recently attracted increasing attention in
computer vision. Generally, the deblurring problem can be represented as
y =k⊗x+n
where y represents the blur image, x is the latent image, k is the blur kernel, n is
Gaussian noise, ⊗ denotes convolution operator. The problem is how to recover
both x and k from single input y, which is highly ill-posed and non-convex.
Most previous works used regularization term method for solving above prob-
lem. The formulation of regularization term method for image deblurring as
J (x, k) = φ (k ⊗ x, y) + λϕ (x, k)
c Springer Nature Singapore Pte Ltd. 2018
B. Huet et al. (Eds.): ICIMCS 2017, CCIS 819, pp. 13–22, 2018.
https://doi.org/10.1007/978-981-10-8530-7_2
14 B. Zhang et al.

where φ (k ⊗ x, y) is data term, ϕ (x, k) denotes regularization constraint term,

and λ are weights.
When n denotes Gaussian noise, the data term
2
φ (k ⊗ x, y) = k ⊗ x − y2
So, many previous methods obtained satisfactory results by carefully designed
ϕ (x, k) and empirically tuned λ. Priors on the statistics of both image and kernel
determine the selection of regularization term. Specifically, many methods focus
on designing effective priors for x, while k is usually restricted to be smooth. So,
priors on the statistics of image are crucial.

Fig. 1. Example of different p value in Lp -norm. The curves in (a) show the solu-
tion space of shrinkage function ming (g − v)2 + λ|g|p , p = 0 (yellow), p = 2/3
(red), p = 1 (black). (c) − (e) represents the solution of the image denoising equa-
tion minx x − y22 + λ ∇xpp . (b) Blurry image, (c) p = 0, (d) p = 2/3 , (e) p = 1.
The regularizer with smaller p value smooth many small-scale textures, larger p value
retain more details. (Color figure online)

In this paper, we propose a blind image deblurring method with adaptive

priors. We introduce a family of priors (Lp -regularized gradient priors) as sparse
regularization. We select different priors along iterations. Different priors cor-
responds to different energy functions. And we minimize them via our mod-
ified ADMM1 algorithm. We will prove the convergence of our algorithm for
any Lp -norm. In other words, we can select any Lp -norm along with iterations.
Specifically, at the initial stage, we select L0 -norm, which can be used to smooth
small-scale textures and select salient edges, so that we can estimate coarse
shape of blur kernel rapidly. Then we refine the estimated blur kernel by gradu-
ally increasing the p value of Lp -norm along with iteration. Finally, we model the
heavy-tailed distribution of gradients in natural scenes via Lp -norm (p = 0.8).
Figure 1 shows one example. The regularizer with smaller p value can smooth
many small-scale textures, and with the increase of p value, the intermediate
images include more gradient details.
The contributions of this work are as follows: (1) we proposed a method
for blind image deblurring with adaptive priors. Specifically it is adaptive Lp -
regularized gradient priors; (2) we proved the convergence of the proposed algo-
rithm; (3) our approach achieved competitive results on widely-used natural
image deblurring benchmarks [7,9,13] and some challenging images.
1
Alternating direction method of multiplier.
 Blind Image Deblurring Using Adaptive Priors 15

2 Related Work
Most image deblurring methods are based on Bayesian theory, which can be
divided into two classes: variational Bayes (VB)-based and maximum a posterior
(MAP)-based. VB-based is more robust, but computationally inefficient.
In recent years, many MAP-based approaches are developed and achieved
promising performance. The priors for MAP-based is critical. In [2] Chan and
Wong used total variation (TV) regularizer for blind image deblurring. Xu et al.
[16] used the L0 -regularizer on the image gradient. Krishnan et al. [8] introduced
L1 /L2 in image gradient prior. However, these approaches failed to address spe-
cific images such as face and low-illumination images.
Different from the method of iteration wise priors [18], which learned iteration
wise parameters of the hyper Laplacian prior for kernel estimation. Our method
design adaptive priors for both kernel and latent image estimation, and we are
not only limited to use hyper Laplace (0 < p < 1) priors but also select Lp -norm
where p ∈ {[0, 1] ∪ 2}. Furthermore, our method can be directly applied to any
blurry images while do not need to train. [11,17] are some other related works.

3 Blind Image Deblurring with Adaptive Priors

In this section, we describe our algorithm and analyze its convergence. Firstly
we prove the convergence of modified ADMM algorithm with two assumptions.
Then we apply the modified ADMM algorithm to image deblurring problems.
And we prove that for any Lp -norm (p ∈ {[0, 1] ∪ 2}) satisfies the assumptions.

3.1 Our Algorithm Framework

Model Statement. We describe our algorithm by solving a widely-concerned
problem. We consider optimization problem:

x̂ = arg min f (x) + λg (x) (1)

x

where f (x) has second order continuous derivative on R. General, g (x) is non-
convex. It is difficult to solve the problem directly. We introduce an auxiliary
variable v, and convert Eq. (1) into a constrained problem:

(x̂, v̂) = arg min f (x) + λg (v) , subject to x = v (2)

x,v

General, we obtain the solution for Eq. (2) by alternatively solving

 2
min f (x) + ρ2t x − x̃(t)  ,
x  2 (3)
min λg (v) + ρt v − ṽ (t) 
2
v

where x̃(t) = ṽ (t) − u(t) , ṽ = x

(t) (t+1)
+ u(t) is known as ADMM. And u(t+1) =
u + x
(t) (t+1)
−v (t+1)
, ρt+1 = γρt (γ > 1) are updated.
16 B. Zhang et al.

Convergence. To ensure global convergence, one sufficient condition is that

g (v) is convex. But in many cases, g (v) is a complex nonsmooth and even
nonconvex function. We assume a mild condition for g (v) and expect a weaker
form of convergence. Before providing the convergence of our algorithm, we first
give some essential assumptions:
n
1. f : [0, 1] → R has bounded gradients, i.e., ∇f (x) ≤ C0 , for some universal
constant C0 . 
    2
2. v (t+1) − ṽ (t)  = Gσt ṽ (t) − ṽ (t)  ≤ λM
2
ρt , for some universal, nonnega-
 (t)   2
tive constant M . Where Gσt ṽ = min g (v) + ρ2t v − ṽ (t)  for updating
 v
v in Eq. (3) and σt = λ/ρt . Under two assumptions, the iterates of our
algorithm in Eq. (3) is convergent.
Proof. Comes from the assumptions, the iterative sequence generated by our
algorithm have the following estimations:
    
 (t+1) (t)  1 C0
x −x ≤ √ 5 C1 + √ (4)
ρt−2 ρ0
   
 (t+1)  1 C0
v − v (t)  ≤ √ C1 + √ (5)
ρt−2 ρ0
  √
 (t+1) (t)  2 C1
u −u ≤ √ (6)
ρt−2
So, for any t we have
 (t+1)     
x − x(t)  + v (t+1) − v (t)  + u(t+1) − u(t) 
 √  t−2
≤ √1ρ0 8 C1 + √Cρ00 1
γ
(7)
 t−2
=Cδ
 √ 
C0
where the notation C  = √1ρ0 8 C1 + √
ρ0 and δ = 1
γ (0 < δ < 1). From
any n < t ≤ N , we have that
N  (t+1)     
x − x(t)  + v (t+1) − v (t)  + u(t+1) − u(t) 
t=n+1  (8)
≤ C  δ n−1 + δ n + · · · + δ N −2
N −n
= C  δ n−1 1−δ
1−δ

When n → ∞, the sequence tends to 0. Hence the sequence generated by our

algorithm is a Cauchy sequence.

3.2 Image Deblurring Algorithm

Different priors corresponds to different energy functions. So we will minimize a
series of energy functions along iterations. For each energy function, we obtain
the solution by estimating sharp latent image and blur kernel, alternately. For
estimating sharp latent image, we use our modified ADMM algorithm.
 Blind Image Deblurring Using Adaptive Priors 17

Models. Our adaptive image deblurring algorithm is carried out using multi-
scale (S, S − 1, . . . , s, . . . , 1). At each scale, we add specific prior Ps (x) to a
standard formulation for image deblurring as
2 2
min k ⊗ x − y2 + ω k2 + λPs (x) (9)
x,k

where x and y denote the latent and blurred images, respectively; k is a blur
2
kernel with the convolution operator ⊗ and L2 regularized term k2 ; γ and ω
are weights. Specific priors Ps (x) are Lp -regularized gradient priors
p
Ps (x) = ∇xpss (10)

Inner Iterations for Estimating Image. We update intermediate images via

2
our algorithm. We define f (x) = k ⊗ x − y2 as data term, and g (v) = Ps (v)
as prior term. And the iteration form
1 2 ρt 

2

x(t+1) = arg min k ⊗ x − y2 + x − x̃(t)  (11)
x 2 2 2

v (t+1) = Gσt ṽ (t) (12)

u(t+1) = u(t) + x(t+1) − v (t+1) (13)

where x̃(t) = v (t) − u(t) , ṽ (t) = x(t+1) + u(t) and σt = λ/ρt . ρt+1 = γρt as a
continuation scheme.

Fig. 2. Convergence of our algorithm. (a) is input image and kernel, (b) is the result
of our algorithm, and (c) is kernel similarity plot. With the increase of iterations, the
kernel estimation is better.

Equation (11) can be regarded as an inversion step as it involves the for-

ward imaging model f (x), and it can be efficiently computed by Fast Fourier
transforms (FFTs).
Note, f (x) is required to satisfy the first assumption. We prove that f (x) =
2
k ⊗ x − y2 has bounded gradients.
Proof. For K ∈ Rn×n with eigenvalues bounded between 0 and 1. The gradient
of f is ∇f (x) = 2K T (Kx − y) and
2
∇f (x) ≤ 2λmax (K) (x + y)
Another random document with
no related content on Scribd:
clever; the Gordian knot of every question might easily be solved in
this way.
Bearing in mind that the question we are now trying to solve is
this, “What is the evidence afforded by Geology, as to the history of
creation, and in what way does the geological age of the world affect
the supposed statement of Scripture, that the world is only 6,000
years old?” I reply thus, and I prefer to use the words of others rather
than my own, lest it should be supposed that I am introducing mere
novelties of opinion on this subject:—“That the first sentence in
Genesis is a simple, independent, all-comprehending axiom to this
effect; that matter, elementary or combined, aggregated only or
organized, and dependent, sentient and intellectual beings have not
existed from eternity, either in self-continuity or in succession: but
had a beginning; that their beginning took place by the all-powerful
will of One Being, the self-existent, independent and infinite in all
perfection, and that the date of that beginning is not made known.”
These are the words of Dr. Pye Smith,[123] of whose name as an
authority, both in matters of science and philology, no one need be
ashamed.
Dr. Redford says, “We ought to understand Moses as saying,
indefinitely, far back, and concealed from us in the mystery of eternal
ages, prior to the first moment of mundane time, ‘God created the
heavens and the earth.’”
“My firm persuasion is,” says Dr. Harris, “that the first verse of
Genesis was designed by the Divine Spirit to announce the absolute
origination of the material universe by the Almighty Creator, and
that it is so understood in the other parts of Holy Writ; that, passing
by an indefinite interval, the second verse describes the state of our
planet immediately prior to the Adamic creation; and that the third
verse begins the account of the six days’ work.”
Dr. Davidson, in his “Sacred Hermeneutics,” says,—“If I am
reminded, in a tone of animadversion, that I am making science, in
this instance, the interpreter of Scripture, my reply is, that I am
simply making the works of God illustrate His word in a department
in which they speak with a distinct and authoritative voice; that it is
all the same whether our geological or theological investigations have
been prior, if we have not forced the one into accordance with the
other. And it may be deserving consideration whether or not the
conduct of those is not open to just animadversion who first
undertake to pronounce on the meaning of a passage of Scripture,
irrespective of all appropriate evidence, and who then, when that
evidence is explored and produced, insist on their à priori
interpretation as the only true one.”
But I quote no more: such are some of the eminent theological
contributions to this department of science:—satisfactory in this
respect, that a fair interpretation of Scripture does not require us to
fix any precise date, much less the inconsiderable one of six
thousand years, as the period of the earth’s formation.
Geology teaches the same thing.—Of the various formations that
compose the earth’s trust, to the ascertained extent of ten miles,
suppose we select two,—the Old Red Sandstone and the Chalk
formations. Laborious and scientific men have been at the pains to
calculate the gradual increase of some of these now proceeding
deposits,—such as the Deltas, in course of formation at the mouth of
the Nile, and at the gorges of the Ganges; and they find that the
progress of the depth of increase is exceedingly small,—probably not
more than a foot in many years. Mr. Maculloch, a name standing
very high for accurate investigation, states, from his own
observation, that a particular Scottish lake does not form its deposit
at the bottom, and hence raise its level, at the rate of more than half-
a-foot in a century; and he observes, that the country surrounding
that lake presents a vertical depth of far more than 3,000 feet, in the
single series of the Old Red Sandstone formation; and no sound
geologist, he hence concludes, will, therefore, accuse the computer of
exceeding, if, upon the same ratio as the contiguous lake, he allows
600,000 years for the production of this series of rock alone.
A last instance which may here be adduced, of the apparent length
of time required for the construction of a particular rock, offers itself
in the Chalk formation. The enormous masses of this rock,
presenting their tall white precipices in such simple grandeur to our
view, might well excite our astonishment at the periods which would
seem needful for their collection and deposition, even if they were
mere inorganic concretions of calcareous matter. But what shall we
say when the investigations of the microscope have lately revealed to
us that these mountains of chalk, instead of being formed of mere
inert matter, are, on the contrary, mighty congeries of decayed
animal life,—the white apparent particles, of which the chalk masses
are composed, being each grain a well-defined organized being, in
form still so perfect, their shells so entire, and all their characteristics
so discoverable, as to cause no doubt to naturalists as to the species
in the animal economy to which they belonged. How justly does Sir
Charles Lyell, who in his “Elements” records at length this surprising
discovery, exclaim,—
“The dust we tread upon was once alive!”

“Look at the lofty precipices which lay naked a slight section of the
Chalk at the Culvers, or the Needles in the Isle of Wight, or the still
loftier Shakspeare Cliff at Dover, and let the mind form a conception,
if it can, of the countless generations of these minutest of living
creatures it must have required to build up, from their decayed
bodies and their shelly exuviæ, layer on layer, those towering masses
thus brought to our view. Who shall dare to compute the time for this
entire elaboration? The contemplation almost advances us a step
towards forming a conception of infinitude.”[124]
I need not dwell longer on the antiquity of the globe:—Geology and
Scripture present no conflicting testimonies on this subject. Our
interpretation of Scripture has, undoubtedly, been modified; but the
living Word itself abideth, in all its grandeur and purity, for ever.
And “the time is not far distant when the high antiquity of the globe
will be regarded as no more opposed to the Bible than the earth’s
revolution round the sun and on its axis. Soon shall the horizon,
where geology and revelation meet, be cleared of every cloud, and
present only an unbroken and magnificent circle of truth.”[125]
The reader shall not be detained so long on the second point of
inquiry, which is
II. “Was death introduced into the world before the fall of man?
and if it was, how are the statements of Scripture, on this question, to
be explained?” To this I have replied by anticipation, that, in my
opinion, death, upon a most extensive scale, prevailed upon the
earth, and in the waters that are under the earth, countless ages
before the creation of man. Into the proof of this position allow me to
go very briefly, although I am well aware that I run the risk of
incurring the charge of heterodoxy, when I state my full conviction,
that death, as well as the world, was pre-Adamite. The general
impression is the contrary; but general impressions are not always
right:—“general impression” is a very unsubstantial ghost to deal
with, very like that cant phrase we spoke of at the beginning of this
lecture,—“the intelligence of the age.” “General impression” has it,
that death was not pre-Adamite; that there was no death before the
fall; and that, to say the contrary, is, at least, to tread on very
dangerous ground. In vain does Geology—“now happily a true
science, founded on facts, and reduced to the dominion of definite
laws”—lay bare the Silurian rocks, and discover even there extinct
forms of life in exquisitely beautiful preservation. In vain does
Geology, after showing us the fossil trilobite and coral, unfold the
volume of the Old Red Sandstone, and show us there the fossil
remains of fish—so perfect that we might imagine them casts rather
than fossils. In vain does Geology open its vast Oolitic system, and
show us there other forms of extinct life in fossil insects, tortoises,
mighty saurians, and huge iguanodons. In vain does Geology lay bare
the Chalk, with its marine deposits; and the Tertiary formation, with
its enormous theroid mammalia, far surpassing in size the largest
animals we are acquainted with. In vain are all these fossil remains
exhibited imbedded in the earth; and in vain do we search, amidst all
these, for one fossil remain of man, or one fossil vestige of man’s
works. The easy, the cheap, the unreflective answer is, “Oh! these
things were created there, or else Noah’s flood left them there.”
Of course, we can fall back upon a miracle as having done all this;
but to have recourse to miracles when no miracle is recorded, is just
to shake our faith in that all-inspired testimony, that supernatural
Book, the existence of which is the great miracle of time. But there
are the fossils! How did they come there if the forms of animal life,
once inhabiting those remains, had not previously lived and died?
Created! What? Created fossils? Then why not, when the Almighty
created man, did he not create, at the same time, some skeletons of
man, and place them in the earth, as he put skeletons of trilobites,
fishes, reptiles, and mammals there? Our common sense and
reverence both reject the idea. As to the puerile notion that Noah’s
flood put them there, did not Noah’s flood overwhelm man as well as
animals? and as the bones of man are as durable as the bones of
animals, how is it that we never meet with a fossil human skull or
thigh bone, or house?
 We believe that death was a part of the divine plan of God’s
creation; that death is a law of all organic life—a necessary law and a
most benevolent provision; that the living structure of all animals
derives its substance from dead organic matter. We believe that,
altogether apart from human sin, preceding and successive
generations must be the order of being; for if there were no death,
animals would soon pass beyond the limit of provision sufficient for
nutritive support, or of localities for suitable habitations. We believe
that if there had been no death prior to man’s sin, it would involve
the supposition that all animals were herbivorous; whereas, even the
little ladybird cannot live without its meal of aphides; and, so
believing, we find our faith in Scripture deepened when, seeing on
every hand the extensive proofs of death, we find man, the moment
he lost his lordship and proud eminence, and reduced himself
voluntarily to the condition of animalism, immediately brought
penally within the influence of that law of death, whose existence he
must have recognised in the death of animals from the first day of his
creation.
Does any one reply, “This is contrary to Scripture?” I ask them
what Scripture teaches that the death of animals is the result of
man’s sin?—rather would not Scripture sanction the thought that
death was a part of the divine plan of God’s creation, and that the
certainty of man’s transgression was the reason for giving this
constitution to nature? True, Milton sings, in his noble poem, that
will live as long as the English language lives—
“Of man’s first disobedience, and the fruit
Of that forbidden tree, whose mortal taste
Brought death into the world, and all our woe:”

but we are not obliged to call the Paradise Lost our Bible; or to quote
Milton as a physiological authority, although the prevalence of the
opinion that death was not pre-Adamite, and a good deal of theology
besides, is more of Miltonic than of Scripture teaching.
I leave this branch of my subject far before it is exhausted: so far
from that, each of the three points enumerated might easily be
expanded into a lecture; and I can only hope that my brevity in
treating these topics will not be misconstrued into a desire to shirk
any of the difficulties with which their investigation is surrounded.
 III. I come, lastly, to the question of the Noachian Deluge, and
shall again repeat my own words: “What was the character of the
Noachian Deluge?—was it partial or universal? and what are the
apparent discrepancies, in this case, between science and the Bible?”
And I have added to this my belief that the Noachian Deluge was
quite partial in its character, and very temporary in its duration: that
it destroyed only those animals that were found in those parts of the
earth habitable by man, and that it has not left a single shell or fossil,
or any drift boulders or pebbles, or any other remains that may be
traced to its action.
Very briefly we shall try and prove this; and perhaps the most
popular way will be the best remembered,—only that the reader will
bear in mind that this little book does not pretend to exhaust the
subject, but only to realize the idea expressed at the beginning of this
chapter. Presuming, that all have in their recollection the Scriptural
account of the Noachian Deluge, instead of quoting words with
which all are familiar, I will only remark, as the basis of my
illustrations, that rain descended, and probably the ocean
overflowed, for forty days; that the waters lay upon the land, and
covered them one hundred and fifty days; that at the end of that time
they began to subside, and that in twelve months and twenty-seven
days they were gone from the face of the earth, and the Noachian
family liberated from the ark.
The question is, was this flood universal, and were all kinds of
animals preserved in the ark? To which my answer, as involving my
belief, is this, that the flood was local, and that only the animals
peculiar to Armenia were provided for in Noah’s ark.
“Oh! but the Bible says it was universal,” says everybody. Yes; but
that, you know, is just the question between us. The terms “all the
earth” seem to imply universality, but they do not necessarily involve
this. “All countries came to Egypt to buy corn;” certainly not all the
world literally, but all the surrounding countries. So there were once
dwelling at Jerusalem devout Jews “from out of every nation under
heaven;” but not literally out of every nation, for the names of the
nations are immediately given, and we find the nations to have been
a few between Egypt and the Black Sea, and between Italy and
Palestine. There are many other illustrations of a similar character:
these will suffice: I only adduce these to show that at the beginning
Scripture does not oblige us to consider “all” as meaning “every one;”
or to understand literally “all the inhabitants of the earth” as
meaning every creature.
Now, looking at the structure and composition of the earth’s crust,
especially its fossiliferous rocks, I am driven to one of three
conclusions, each of them involving difficulty, I acknowledge, but the
one that involves the least is, of course, the most preferable. Either I
must admit—
1. That the fossils in these rocks were all deposited in order and in
succession, without injury, through a crust of rocks ten miles in
thickness, during twelve months’ violent diluvial action:
2. Or that they were all deposited there during the 15,000 or
16,000 years that had elapsed since the creation of man prior to the
Deluge; that is, supposing the creation of man and the creation of the
earth to have been synchronous. Or, lastly, which theory I accept—
3. That the date of the earth’s physical being is unknown to us, and
that the fossiliferous rocks were deposited in decades of ages before
the creation of man.
For, on the other hand, let us suppose the flood to have been
universal, in the strict and literal sense of the term; then let me
suggest some of the consequences and difficulties of such a theory.
1. One consequence would be that some remains of man or of his
works would have been found; but nothing of this kind has occurred.
Even Armenia has been geologically examined, and no human
remains have been found; and surely man’s bones would last as long
as the shells of a trilobite or terebratula?
2. And, secondly, the organic remains, the fossils themselves,
would have been found confusedly heaped together; whereas, the
remains in the crust of the earth are as carefully arranged as the
contents of a well-ordered cabinet. We know always to a certainty
what fossils will be found in any rock before we examine that rock.
3. Besides which, some, at least, of the organic remains found
ought to correspond with existing beings and species: yet the
contrary is the case, except only a few fossils found near the surface
of the earth, in that portion of the earth’s crust occupied by the
tertiary system.
 Nor is this all. Consider the vast difficulties the universal flood
theory has to contend with, all of which are removed by the theory
we have adopted.
1. There is the quantity of water required. If all over the earth the
water rose twenty-two feet six inches above the tops of the highest
mountains, the quantity of water required would be eight times the
whole quantity of water now existing. Where all this could have come
from first and gone to afterwards, are prodigious stumbling-blocks.
Of course we can resort to miracle; but this is not the way to get rid
of difficulty in a manly and honest spirit.
2. Then consider the number of animals the ark must have
contained. There are 1,000 species of mammalia, 5,000 species of
birds, 2,000 species of reptiles, and 120,000 well-ascertained and
distinct species of insects. Do we pretend that all these were housed
and fed for nearly thirteen months in a vessel that was only 450 feet
long, 75 feet broad, and 45 feet high; and that such a vessel contained
room for them, and their food, besides that of man, for such a long
period. The little toys of Noah’s ark are certainly pretty, but very
mischievous, and most of the popular notions of the flood have
grown up from our nurseries as much from the use of this toy in this
case, as from the reading of Paradise Lost in the other: and the result
is, the Bible is made responsible for it all.
3. Then consider the subsequent distribution of animals: the polar
bear and the tropical elephant, the ferocious tiger and a young fawn,
going out together in order, and without violence: of course we can
suppose another miracle to repress passions and violence. Besides
which, in addition to the fauna, the animal kingdom, we must ask
what became of the flora or vegetable kingdom during this period, if
the flood were universal? We have at least twenty-five botanical
provinces, with their peculiar and numberless farms of vegetable life;
what became of them? Were they preserved in the ark, or under the
water?—for such questions must be answered by those who charge us
with inconsistency in attempting to reconcile the facts of science with
the words of Scripture. And as a last difficulty, (suggested first, I
believe, by Dr. Pye Smith, and which I shall therefore state in his
words, lest it should seem that I use “plainness of speech,”) let us
look at the descent from Ararat out of the ark, into Armenia, with all
these animals, birds, insects, plants and trees. “That mountain is
17,000 feet high, and perpetual snow covers about 5,000 feet from
its summit. If the water rose, at its liquid temperature, so as to
overflow that summit, the snows and icy masses would be melted;
and on the retiring of the flood, the exposed mountain would present
its pinnacles and ridges, dreadful precipices of naked rock, adown
which the four men and the four women, and with hardly any
exception the quadrupeds, would have found it utterly impossible to
descend. To provide against this difficulty, to prevent them from
being dashed to pieces, must we again suppose a miracle? Must we
conceive of the human beings and the animals as transported
through the air to the more level regions below; or that, by a miracle
equally grand, they were enabled to glide unhurt adown the wet and
slippery faces of the rock?”
Such are some of the difficulties and some of the consequences
that must flow from an acceptance of any other theory than the one I
have proposed: that the flood was partial in its character, extending
only over the habitable parts of the earth; and that it was so
temporary in its character as not to have left a single trace of its
influence visible on rock or fossil.
I have thus endeavoured to suggest points of reconciliation
between the accepted facts of Geology and the recorded statements
of Scripture; and if this slight contribution be accepted as an aid to
faith, and a proof of candour on my part to meet those who linger on
the border land of doubt, my purpose will be fully answered.
Let me add, in the words of Chenevix Trench—words uttered in the
University of Cambridge not long since: “May we in a troubled time
be helped to feel something of the grandeur of the Scriptures, and so
of the manifold wisdom of that Eternal Spirit by whom it came; and
then petty objections and isolated difficulties, though they were
multiplied as the sands of the sea, will not harass us. For what are
they all to the fact, that for more than 1,000 years the Bible
collectively taken, has gone hand in hand with civilization, science,
law—in short, with the moral and intellectual cultivation of the
species, always supporting and often leading the way? Its very
presence as a believed book, has rendered the nations emphatically a
chosen race; and this, too, in exact proportion as it is more or less
generally studied. Of those nations which in the highest degree enjoy
its influences, it is not too much to affirm that the differences, public
and private, physical, moral, and intellectual, are only less than what
might be expected from a diversity in species. Good and holy men,
and the best and wisest of mankind, the kingly spirits of history
enthroned in the hearts of mighty nations, have borne witness to its
influence, and have declared it to be beyond compare the most
perfect instrument and the only adequate organ of humanity: the
organ and instrument of all the gifts, powers, and tendencies, by
which the individual is privileged to rise beyond himself, to leave
behind and lose his dividual phantom self, in order to find his true
self in that distinctness where no division can be,—in the Eternal I
am, the ever-living Word, of whom all the elect, from the archangel
before the throne to the poor wrestler with the Spirit until the
breaking of day, are but the fainter and still fainter echoes.”

M. CLAY, PRINTER HEAD STREET HILL.

 1. Whewell’s Astronomy and Physics, p. 48.
2. From παλαιός, ancient, and ζωόν, life; ancient-life period.
3. Hughes, Physical Geography. 3d ed. p. 21.
4. Hughes, Physical Geog. p. 22.
5. Dr. Pye Smith.
6. As Chimborazo in South America, 21,414; Ararat, 16,000; Dhawalagiri, in
the Himalayas, 28,000 feet above the level of the sea; compared with which what a
mole-hill is Vesuvius, only 8,947 feet; or Blue Mountain Peak, 8,600, or even Mont
Blanc, that monarch of mountains, which is 15,816 feet above the sea!
7. Hughes, p. 16.
8. Chambers’ Rudiments of Geology, p. 71.
9. These wells are so frequently spoken of as to need no explanation, further
than to remind the reader that they are so called from having been first introduced
in the province of Artois, the ancient Artesium in France.
10. The deepest Artesian well is the famous one in the Plaine de Grenelle,
Paris. This well yields 516 gallons a minute; its temperature is 81° Fahr.; and its
depth is nearly 1,800 feet.
11. How truly hieroglyphics—sacred carvings; (ieros, sacred, glupho, I carve;)
and in this sense there is a holier meaning than Shakspeare could have dreamt of
in his well-known lines, when applied by the geologist to his researches:—

“And this our life, exempt from public haunt,

Finds tongues in trees, books in the running brooks,
Sermons in stones, and good in everything.”

12. And I may say, my friend also, to whom, during my residence in Jamaica, I
was frequently indebted for contributions on natural history to the Jamaica
Friendly Instructor, of which I was Editor.
13. A Naturalist’s Sojourn in Jamaica, by P. H. Gosse, Esq. pp. 496–7.
14. So called because of its grained or granular appearance.
15. First brought from Syene, in Egypt.
16. Feld-spar, written also felspar, a compound of feld, field, and spar.
17. See Ansted’s Ancient World, p. 21.
18. Memnon, or Ramesis. This famous head is in the British Museum; the
body is of greenstone, the head of syenite, and the bust one continuous mass.
19. From dis and integer. The separation of the whole parts of a rock, without
chemical action, by means of the light, the air, or the rain, is called disintegration.
 20. Lieut. Portlock on Geology, p. 93.
21. Ansted’s Geology, Descriptive and Practical, vol. ii. pp. 290, 291.
22. Ansted’s Geology, p. 291.
23. As the ancients did not know or use the compound metal brass, though
bronze was common amongst them, we must in this verse, and all others in which
the word “brass” is used, understand it to mean copper.—Hughes’ Scripture
Geography, Art. Geology of Palestine, p. 133.
24. Murray’s Hand-book for Cornwall, p. 199.
25. Ansted, vol. ii. p. 418.
26. Whewell, Anniversary Address to Geol. Society, 1839.
27. In Memoriam.
28. Dr. Pye Smith says 140,000 feet.
29. See a valuable map of fossils published by the Christian Knowledge
Society.
30. Trilobite: treis, three, and lobos, a lobe; having three lobes.
31. Bridgewater Treatise, vol. i. p. 396.
32. A fossil shell allied to the Argonauta and Carinaria.
33. “Man has no tail, quantum mutatus; but the notion of a much-ridiculed
philosopher of the last century is not altogether without foundation; for the bones
of a caudal extremity exist in an undeveloped state in the os coccygis of the human
subject.” Poor man!—Vestiges of Creation, p. 71.
34. Sedgwick, p. 216, “On the Studies of the University of Cambridge.”
35. “My School and Schoolmasters,” by Hugh Miller.
36. “Old Red Sandstone; or, New Walks in an Old Field;” by Hugh Miller, p.
48.
37. “By mind, by hand, and by hammer.”
38. “Old Red Sandstone,” p. 66.
39. Ichthyolite: ichthus, a fish, lithos, a stone: fossil fish, or the figure or
impression of a fish in the rock.
40. “Old Red Sandstone,” pp. 41, 42.
41. “Old Red Sandstone,” p. 69.
42. From akanthos, a thorn, and pterugion, the fin.
43. From malakos, soft, and pterugion, the fin.
 44. 1. Ganoid, from ganos, splendour, because the scales are coated with a
bright enamel.
45. 2. Placoid, from plax, a plate; sometimes large, sometimes reduced to a
point; e.g. shark.
46. 3. Ctenoid, from kteis (gen. ktenos, a comb); scales jagged like a comb.
47. 4. Cycloid, from kuklos, a circle; scales smooth and simple: e.g. salmon,
&c.
48. From kephalē, the head; aspis, a buckler.
49. Coccosteus, from kokkos, a berry, and osteon, a bone.
50. “Old Red Sandstone,” p. 86.
51. Pterichthys: pteron, a wing, and ichthus, a fish.
52. “Old Red Sandstone,” pp. 80, 81.
53. Osteolepis: osteon, a bone, and lepis, a scale.
54. Operculum, the flap which covers the gill.
55. “Old Red Sandstone,” p. 111.
56. “Vast quantities:” let any reader go and turn over the non-bituminous
shale lying on the waste heaps of every coalpit, and he will see that this is no
exaggeration.
57. Capillus Veneris.
58. Corruption of arrière-dos, a fire-place. See a view and description of one
in “A Visit to Penshurst,” in Howitt’s “Visits to Remarkable Places,” Second Series.
59. Juicy and soft, as peas, beans, plantains, bananas, &c.
60. “Ancient World,” pp. 76, 77.
61. This may seem strange at first; but I have journeyed through tropical
forests that realized completely this sketch, so far as stillness and silence are
concerned. A modern and most accomplished naturalist says of a Jamaica virgin
forest, “Animal life is almost unseen; the solitude is scarcely broken by the voices
of birds, except that now and then the rain-bird or the hunter (large cat-tailed
cuckoos that love the shade) sound their startling rattle, or the mountain partridge
utters those mournful cooings which are like the moans of a dying man.”—Gosse’s
Jamaica, p. 198.
62. From κάλαμος (calamus), a reed.
63. Ansted’s “Ancient World,” p. 82.
64. Mesozoic: i.e. middle life period; mesos, middle, zoos, life.
65. The Religious Tract Society.
 66. Lyell’s “Manual of Elementary Geology.” Postscript, p. 13.
67. Ansted’s Geology, vol i. p. 306.
68. Ichnites; from ichnon, a footstep, and eidos, like.
69. Ornithos, a bird, and ichnon.
70. Marsupial, from marsupium, a pouch; animals of the fourth order of
Cuvier, that have a pouch in which the young are carried.
71. Batrachian, from batrachos, a frog; animals in Cuvier’s fourth class of
reptiles.
72. Cheir, the hand, therion, a wild beast; a wild beast with a foot like a hand.
73. From labyrinthus, a labyrinth, and odous, a tooth; so called from the
labyrinthine structure of the tooth.
74. In some cases we find, corresponding to a set of footmarks, a continuous
furrow, presumed to be the impression of a tail dragged along the sand by the
animal while walking.
75. Ansted’s Ancient World, pp. 125–127.
76. Knight’s Cyclopædia of Arts, &c.
77. Quarterly Review, May, 1852. Article on Roger de Coverley.
78. This is a corruption, we are inclined to think, of the word “layers;” one of
those provincial corruptions of the Queen’s English that get stereotyped.
79. Buckland’s Bridgewater Treatise. pp. 351, 352.
80. A fossil bivalve, allied to the oyster, and very abundant in the secondary
strata.
81. Belemnite, from belemnos, a dart, and so called from its arrow-headed
shape.
82. Saurian, from sauros, a lizard, the name by which the great family of
lizards is designated.
83. From ichthus, a fish, and sauros, a lizard; so called from its resemblance to
both.
84. Heteroclite; heteros, another, and klitos, inclining; a word applied to any
thing or person deviating from common forms.
85. Very unlike the alligator, whose eyes are placed at a considerable distance
behind the nose.
86. From pleiōn, more, and sauros, a lizard; because it is more like a lizard
than the Ichthyosaurus.
87. Mantel’s Fossils of the British Museum, p. 341.
 88. This formation is sometimes called the Jurassic system.
89. Lyell’s Manual of Elementary Geology, p. 12, ed. 1852.
90. “So vast an expanse!” Mr. Darwin traced coral reefs in the Pacific, 4,000
miles long and 600 broad. Between the coasts of Malabar and Madagascar there is
a chain of coral reefs, called the Maldives and Laccadives, 480 miles long and 50
miles wide. On the east coast of Australia there is an unbroken reef of 350 miles
long; and between Australia and Guinea, coral reefs extend 700 miles in length.
Truly the coral animals, like the “conies,” are a “feeble folk,” but their habitations
survive our proudest monuments.
91. Hugh Miller’s First Impressions, pp. 203, 204.
92. Brash is s Wiltshire word for short or brittle; and thus a quick-tempered,
irritable person, is said to have a brashy temper.
93. Geology for Beginners (Weale’s Series), p. 147.
94. Juke’s Popular Geology, pp. 42–44.
95. From krinos, a lily, and eidos, like; lily-shaped animals of the Radiated
division, forming a link between the animal and vegetable world.
96. From trochos, a wheel; wheel-shaped crinoideans.
97. From pteron, a wing, and dactulos, a finger; the wing-fingered animal.
98. The term Weald or Wold is the old Saxon for our present Wood; and now,
altered by pronunciation, is found in connexion with many words and names of
places: e.g. Waltham (Weald-ham), the wood house or home; Walthamstow, the
wood house store, and so on. Thus it is that words are “fossil poetry.”
99. Alison’s description of South America, in History of Europe (Article, South
American Revolution); vol. viii.
100. “Our disposition is, and has been, not to multiply miracles after the sort
in which this has been done by many more zealous than wise friends of revelation.
In all cases we allow the miracle without question, which is distinctly claimed to be
such in the Scriptures, and where the circumstances clearly indicate that a miracle
was necessary,—we say ‘necessary,’ because we are persuaded that the Almighty
has almost invariably chosen to act through natural agencies, and under the laws
which he has imposed on nature, whenever they are adequate to produce the
required result. We believe it is one of the beautiful peculiarities of the Bible, that it
has none of those gratuitous and barren wonders, which form the mass of the
pretended miracles which the various systems of false religion produce.... For our
own part, we do not wish to hear of small miracles, which leave us doubtful
whether there be any miracle at all. If we are to have miracles, let them be
decidedly miraculous, and let not our veneration for the Divine character be
offended by exhibitions of the Almighty, as laying bare his holy arm to remove the
small remaining difficulty which theorists leave him to execute.”—Dr. Kitto’s
Biblical History of Palestine.
 101. We have a fine specimen before us which we brought from Demerara,
answering well to Gosse’s description of the iguana found in Jamaica. “In the
eastern parts of the island the great iguana (Cyclura lophoma), with its dorsal
crest, like the teeth of a saw, running all down its back, may be seen lying out on
the branches of the trees, or playing bo-peep from a hole in the trunk.” It is
considered a great delicacy by many, but it never seemed Christian food to us, and
we never ventured to provoke our palate with a taste.
102. Enaliosaurians are sea lizards, such as those found in the Lias; and
deinosaurians are terrible lizards, such as those found in the Wealden.
103. Ansted’s Ancient World, pp. 164–168.
104. Just published by Bohn, in his valuable “Scientific Library;” a marvel of
cheapness and value.
105. Since writing the above we have met with the following, which proves that
this origin of chalk is not so fabulous as some think it:—“Lieut. Nelson, Mr. Dance,
and others have shown, that the waste and débris derived from coral reefs
produces a substance exactly resembling chalk. I can corroborate this assertion
from my own observations, both on some very white chalky limestones in Java and
the neighbouring islands, which I believe to be nothing else than raised fringing
coral reefs, and on the substance brought up by the lead over some hundreds of
miles in the Indian Archipelago, and along the north-east coast of Australia, and
the coral sea of Flinders,”—Juke’s Physical Geology, p. 263.
106. We take the origin of the word Folkstone to mean, that that old town was
once built of the brick that may be made of the galt: it was the folk’s-stone.
107. “The Religion of Geology,” &c., by E. Hitchcock, LL.D. &c. p. 70.
108. Mantell’s “Geological Excursions,” p. 145.
109. Richardson, p. 391.
110. Under-borne rocks; upo, below, and ginomai, to be formed.
111. Middle life period: mesos, middle, and zoos, life.
112. Recent-life period: kainŏs, recent, and zoos, life.
113. Juke’s Practical Geology, p. 265.
114. Lyell’s Manual of Geology, pp. 97, 98.
115. Ansted’s Geology, vol. ii. p. 14.
116. Even Hitchcock’s good book is sadly disfigured and damaged, by trying to
make geology prove too much. How can geology teach or suggest sin and the
resurrection?
117. British Quarterly, Feb. 1852.
118. Owen’s British Fossil Mammals and Birds, p. 255.
 119. Mantell, pp. 477–479.
120. Mantell, p. 471.
121. The skeleton is not more than 150 years old, and is probably one of an
Indian who fell in war; and has been covered with carbonate of lime, held in
solution in some spring.
122. Hugh Miller’s “First Impressions of England and its People,” p. 362.
123. No sooner did geology give signs of being able to speak from her
subterranean abode, and say something new about the history of this old world,
than Dr. Smith was among the foremost of the geologists, intent upon the
interpretation of these mysterious, and at first incoherent sounds. At times the
sounds seemed unscriptural, but his faith never failed; at other times it seemed in
confirmation of Scripture, and he was filled with delight. There were sepulchres
older than what he had accounted the era of death, and he must solve the mystery.
Mineralogy, to which from his youth he had given considerable attention, became
to him history more ancient than that of Moses, and poetry more fascinating than
that of Homer. His minerals became books of wonderful tales; his fossils, before
riddles of nature, the pictures of things in ancient worlds. The earth was a land of
monuments, and the rock which before seemed nothing more than the solid
masonry of the foundation on which men might build their dwellings, became the
enduring chronicle of the millions of years in which extinct ages had risen,
flourished and decayed. From that time he suffered no discovery of the geologists
to escape his attention; and every valuable book upon the subject in English,
German, or French, contributed its supplies to mitigate his insatiate craving after
further information.
Dr. Smith had another reason for devoting a large proportion of his time to
geological studies. The new science had something to say about Holy Scripture. It
threatened, as many understood its first ambiguous words, to contradict the book
of Genesis.
Whatever affected theology was of supreme importance in the estimation of
the Homerton professor. Having full confidence in the truth of God’s word, he was
sure that nature and revelation, however they appeared to superficial observers,
could not be really at variance. In that confidence he patiently listened to every
word the new science had to say about the creation of the world. To him belongs
the honour, in the opinion of the most eminent geologists, of having relieved their
science of every appearance of hostility to Scripture. Of his book on this subject,
Dr. Mantell said, “It is, indeed, the dove sent out from the ark of modern geology;
and it has returned with the olive branch in its mouth.”—British Quarterly, Jan.
1854. Art. Dr. Pye Smith.
124. Gray’s Antiquity of the Globe, pp. 57–59.
125. Hitchcock, p. 70.
 August, 1854.

A
GENERAL CATALOGUE
OF WORKS PUBLISHED BY

ARTHUR HALL, VIRTUE & CO.

25, PATERNOSTER ROW, LONDON.
ACKWORTH VOCABULARY, or English Spelling Book; with the
Meaning attached to each Word. Compiled for the use of
Ackworth School. New Edition. 18mo. cloth, 1s. 6d.
AKERMAN (J. Y.),—LEGENDS OF OLD LONDON. Post 8vo.
cloth, gilt, 3s. 6d.
ALFRED DUDLEY; or, the Australian Settlers. Second Edition.
With Nine Illustrations. 16mo. cloth, 2s. 6d.
ANALYSIS AND SUMMARY OF OLD TESTAMENT HISTORY
AND THE LAWS OF MOSES, with a connexion between the
Old and New Testaments, an Introductory Outline of the
Geography, Political History, &c. By J. T. Wheeler, F.R.G.S.
Fourth Edition, post 8vo. cloth, 5s. 6d.
—— NEW TESTAMENT HISTORY; including, 1. The Four Gospels
harmonized into one continuous Narrative. 2. The Acts of the
Apostles, and continuous History of St. Paul. 3. An Analysis of
the Epistles and Book of Revelation. 4. An Introductory Outline
of the Geography, Critical History, Authenticity, Credibility,
and Inspiration of the New Testament. The whole Illustrated
by copious Historical, Geographical, and Antiquarian Notes,
Chronological Tables, &c. Second Edition, revised. Post 8vo,
cloth, 5s. 6d.
—— A POPULAR ABRIDGMENT OF OLD AND NEW
TESTAMENT HISTORY, for Schools, Families, and General
Reading, Explained by Historical and Geographical
 Illustrations, and numerous Map Diagrams. Two Vols. 18mo.
cloth. 2s. 6d. each.
—— THE HISTORICAL GEOGRAPHY OF THE OLD AND NEW
TESTAMENTS. Illustrated with Five coloured Maps, and large
View of Jerusalem, with a Plan of the Ancient City. Folio, cloth,
7s. 6d.
ANATOMY AND PHYSIOLOGY OF THE HUMAN MIND,
(MANUAL OF THE.) By Rev. J. Carlile, D.D. Fcap. cloth, 5s.
APOCALYPTIC SKETCHES; or, LECTURES on the BOOK of
REVELATION. Delivered in Exeter Hall, and at Crown Court
Church. By the Rev. John Cumming, D.D. New Editions,
revised and corrected, with Indices. Sixteenth Thousand.
Three Vols. Fcap. cloth, gilt, 9s. each; morocco extra, 39s.
ART JOURNAL (The), an Illustrated Journal of the Fine Arts, the
Arts of Design, Manufacture, &c. &c. Published Monthly, with
Two Engravings from the Vernon Gallery, and between Forty
and Fifty fine Engravings on Wood. 2s. 6d.
—— New Series, Volumes I. to V. Each with Twenty-four Line
Engravings from the “Vernon Gallery,” Twelve Engravings of
Statues on steel, and about Eight Hundred Engravings on
Wood. Super-royal 4to. cloth, 1l. 11s. 6d. each.
—— ILLUSTRATED CATALOGUE of the Great Industrial
Exhibition of 1851, containing upwards of Fourteen Hundred
Engravings on Wood, and a Frontispiece on Steel, Imperial 4to.
cloth, gilt edges, 21s.
—— of the Great Exhibition in Dublin, with numerous Engravings.
Cloth, gilt edges, 10s.
ART OF CHESS-PLAY. A NEW TREATISE ON THE GAME OF
CHESS. By George Walker, Esq. Fourth Edition, 12mo. cloth,
2s. 6d.
AUTOBIOGRAPHY OF WILLIAM JERDAN; With his Literary,
Political, and Social Reminiscences and Correspondence,
during the last Forty Years, as Editor of the “Sun” Newspaper,
1812–17, and of the “Literary Gazette,” 1817–50, in connexion
with most of the Eminent Persons who have been distinguished
 in the past half-century as Statesmen, Poets, Authors, Men of
Science, Artists, &c.; with Portraits, &c. Now complete. In 4
Vols. post 8vo. cloth, 21s.
BAKEWELL’S (Frederick C.) PHILOSOPHICAL
CONVERSATIONS, in which are familiarly explained the
causes of many daily occurring Natural Phenomena. Third
Edition, with Cuts. Fcap. cloth, 3s. 6d.
BALLADS FOR THE TIMES, now first collected. American Lyrics,
Geraldine, Modern Pyramid, Hactenus, A Thousand Lines, and
Other Poems, by Martin F. Tupper, D.C.L. F.R.S. Third
Edition, with Vignette and Frontispiece, uniform with
“Proverbial Philosophy,” fcap. cloth, 7s. 6d.
“With smoothness of measure, Mr. Tupper’s design is always
excellent, and his versification is brought to bear upon things
of no transient interest. It is one of the best characteristics of
his labours, that he does not write for praise, but for the
benefit of his fellow-men—not merely for time, but for
eternity.”—Bell’s Messenger.
BAPTISMAL FONT (The), an Exposition of the Nature and
Obligations of Christian Baptism. With an Appendix. By the
Rev. John Cumming, D.D. Fourth Edition, fcap. cloth, gilt
edges, 2s.
BARBAULD’S (Mrs.), LEÇONS POUR DES ENFANS, depuis l’âge
de Deux Ans junqu’à ’Cinq. Avec une Interprétation Anglais.
New Edition, 18mo. cloth, 2s.
BARTLETT (W. H.),—FOOTSTEPS OF OUR LORD AND HIS
APOSTLES, in Syria, Greece, and Italy. A succession of Visits to
the Scenes of New Testament Narrative. With Twenty-three
Steel Engravings, and several Woodcuts. Third Edition, super-
royal 8vo. cloth, gilt edges, 14s.; morocco elegant, 26s.
—— FORTY DAYS IN THE DESERT, on the Track of the Israelites;
or, a Journey from Cairo by Wady Feiran to Mount Sinai and
Petra, with Twenty-seven Engravings on Steel, a Map, and
numerous Woodcuts. Fifth Edition, super-royal 8vo. cloth full
gilt, morocco elegant, 21 s.