search

sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)
MIT License
2.42k stars 668 forks source link

question about your model #48

Closed OValery16 closed 6 years ago

OValery16 commented 6 years ago

I read your paper and I go deeply impressed. I wanted to know how would you compare your work to GAN-oriented approach such as SRGAN (advantage/drawback)

Thanks,

sanghyun-son commented 6 years ago

Hello.

The major drawback of our method is model size.

Since we got enhanced performances by increasing the model size, we have to further optimize our method to use less resources.

The advantage is that we do not use batch-normalization, which is a quite expensive operation.

Our EDSR-baseline model can perform better than SRResNet (from SRGAN) even with out batch-normalization.

Also, we validated various techniques for super-resolution (self-ensemble or multi-scale learning), which can be further researched in future.

Thank you.

OValery16 commented 6 years ago

Thank for you for your very clear explanation