You are on page 1of 3

Critical Review

Fareena 19L-1804 Talha 19L-1871


December 6, 2020

Sentiment Prediction by Merging Hand-Craft and


CNN Features

1
1. Citation
[1] . Wang Fengjiao and Masaki Aono Toyohashi University of Tech-nology
Toyohashi, JAPAN

2. Area
Visual Sentiment Analysis

3. Keywords
Visual Sentiment; Convolutional Neural Networks; Hand-craft Fea-tures;
Data Augmentation; Transfer Learning

4. Problem Statement / Hypothesis


How to predict visual sentiment by merging hand-craft and CNN features

5. Propsed Solution
They plan to merge CNN which has shown remarkable achievements in
Computer Vision recently, with handcraft features like Color His-togram (CH)
and Bag-ofVisual Words (BoVW) with some local fea-tures like SURF and
SIFT to predict sentiment of images. The results of experiments demonstrate
that our proposed visual sentimental pre-diction framework outperforms the
present state-of-the-art methods Proposed a novel framework considering
both hand-craft and CNN features to predict sentiment of images

6. Major results
By merging hand-craft and CNN features can improve performance of visual
sentiment prediction. Smart approaches to data augmenta-tion can increase
the size of the training set 10 times or more

Method : DA : NDA

3 Agree Accuracy : 0.744 : 0.720

4 Agree Accuracy : 0.866 : 0.828

= 5 Agree Accuracy : 0.921 : 0.853

7. Significance

2
1. Sentiment analysis of image data and Convolutional Neural Net-work
(CNN) features

2. Both hand-craft and CNN features to predict sentiment of images

3. Employ both data augmentation and transfer learning from a pre-trained


CNN such as VGG16 trained with ImageNet dataset.

4. proposed visual sentimental prediction framework outperforms the


current state-of-the-art methods

8. Shortcomings
1. Polarity sentiment classification was limited

2. Experimental details are not defined

3. Hardware specification details are not defined

4. The research on sentiment analysis of visual contents has not been


deeply investigated yet

5. select only a 64-bin color histogram feature

9. Future Directions
1. Only focused on polarity sentiment classification, while there is
potentially more sentiment such as anger, sad, and surprise.

2. Each image may have multiple sentiment, which reduces to “multi-label


multi-class” problems.

3. Finally, even one emotion may have several intensity levels such as
“strong positive” and “weak positive”, where we might need “fine grained”
sentiment analysis with more sophisticated framework

10. Any cited reference to follow up


1. C. Szegedy, W. Liu, Y. Jia and P Sermanet, “Going deeper with
convolutions,” Proceedings of the IEEE conference on computer vi-sion and
pattern recognition, 2015, pp. 1-9.

11. Any comments


NA

You might also like