You are on page 1of 1

IMAGE RECOGNITION, SEMANTIC SEGMENTATION AND PHOTO

ADJUSTMENT USING DEEP NEURAL NETWORKS

BY

ZHICHENG YAN

Key point written in introduction

1. In this thesis author described a collection of DNN-Based module to address three challenging
computer vision tasks, namely large-scale visual recognition, image semantic segmentation and
automatic photo adjustment.

2. Author introduces method of image recognition, segmentation and adjustment by using latest
technology DNN which has MLP, CNN and RNN feature for better vision purpose.
3. The previous study shows that CNN used to recognize digits and characters in documents and
gradually researchers study observed the effective GPU implementation is available in CNN due
to presence of largescale ImageNet database.
4. 80K lemmas can be populated by ImageNet with a high resolution than only 600 images can be
seen by IFAR100.
5. This study also reveals that the difficulty is seen in object separation from image which can be
determined by N-way Flat classifiers for discriminating among categories.
6. Author have developed a hybrid net, which consist of alternating convolutional layers and
spatially recurrent layers.
7. Few attempts were made to hieratical classification along with training algorithm which
deteriorating image concede can conclude HD-CNN results less error than CNN.
8. Modifying CNN to FCNN can segment image by discriminated pixel wise feature.
9. Recent study tells that RNN model increases long range dependency and image factors.
10. Assuming combining usage of CNN and RNN may produce best aspect.
11. RNN and CNN works on raw image pixel whereas MLP finds fit for photo adjustment activities.
Common practice is seen in various camera usage apps like Instagram.
12. MLP Network is trained to develop large number of super pixels can be obtained.
13. The resulting model can be deemed as an ensemble of specialized classifiers.
14. Results can be achieved on datasets with both global and local stylized adjustments.

Syed Khalid Mehmood (28987)

You might also like