Hello all
I want to employ deep learning technology in my project which is about super resolution .. can anyone inform me in general the advantages and limitations of GAN model .. Especially, i want to see the opinions of the persons who worked practically on that model on different applications..Regards
If you take a closer look at reconstructed image of SRGAN, without seeing the ground truth image, it looks quite realistic. But if you compare the original image together with the recovered SRGAN image, the texture is not recovered correctly. GAN fills something else to the texture part which does not exist in the original picture.
The observation becomes stronger in the 'text' regions. If the text is too small or distorted in the original image, SRGAN recovers totally different things. It is no longer text or readable in SRGAN. It is because of the loss function I guess. It makes the image look realistic in general --> leads to favor natural scene but not unnatural one (text, artificial structure). Same thing applied for super-resolution of human face. It can result totally different person.
This is somehow dangerous if you perform classification task related to these region (i.e. texture classification, licence plate recognition, etc.)
Sorry that I cannot attach the picture but you can take the closer look at visual result in the paper.
Best regards,
I am also begin working with GAN, and from the point of view of conventional method guy the pros and cons of GAN are following (I will update later along with my GAN knowledge)
- Pros:
+ very impressive result related to content genrated, and realistic visual content.
+ also can be modeled for other applications like image restoration
- Cons:
+ Difficult for training: huge computation
+ "realistic" but not real. Fake pattern can be created, especially for small scale structure or non-nature object like text ---> problem of loss function.
Some physical effects in picture like shadow will become unrealistic.
Many thanks Thuong Nguyen Canh ..Can you more illustrate this please
( + "realistic" but not real. Fake pattern can be created, especially for small scale structure or non-nature object like text ---> problem of loss function.)
I will wait your updates ..Regards
If you take a closer look at reconstructed image of SRGAN, without seeing the ground truth image, it looks quite realistic. But if you compare the original image together with the recovered SRGAN image, the texture is not recovered correctly. GAN fills something else to the texture part which does not exist in the original picture.
The observation becomes stronger in the 'text' regions. If the text is too small or distorted in the original image, SRGAN recovers totally different things. It is no longer text or readable in SRGAN. It is because of the loss function I guess. It makes the image look realistic in general --> leads to favor natural scene but not unnatural one (text, artificial structure). Same thing applied for super-resolution of human face. It can result totally different person.
This is somehow dangerous if you perform classification task related to these region (i.e. texture classification, licence plate recognition, etc.)
Sorry that I cannot attach the picture but you can take the closer look at visual result in the paper.
Best regards,
The effect of GAN on super-resolution is well studied in the following works. Have a look at them.
The perception-distortion tradeoff, CVPR 2018 - https://arxiv.org/pdf/1711.06077.pdf
2018 PIRM Challenge on Perceptual Image Super-resolution - https://arxiv.org/pdf/1809.07517.pdf
+
the works that got published under the same challenge which can be accessed from https://pirm2018.org/papers.html
1. GAN are good classifier and data synthehsizer than CNN
2. GAN can works well with lack of data but CNN are data hungry
3. GAN are state of the art models
4. GAN are good for Segmentaion and classifcation but CNN are the best for classification (If dataset is large)
- Badges
- Science topic