2016_super_resolution

Image Super-Resolution Using Deep Convolutional Networks ICCV2015

I re-implement this paper and includes my train and test code in this repository. This code uses MIT License.

Note that:

Thanks for @star4s. I fixed some bugs in the network training code and made the code more clear to use. (2017/4/29)

Training data

I random selected about 60,000 pic from 2014 ILSVR2014_train (only academic) You can download from (Sorry. My Google Driver is out of storage. So I remove it. ) or BaiduYun

Result

This code get the better performance than 'bicubic' for enlarging a 2x pic. It can be trained and tested now.

original pic -> super resolution pic (trained by matconvnet)

How to train & test

1.You may compile matconvnet first by running gpu_compile.m (you need to change some setting in it)

For more compile information, you can learn it from www.vlfeat.org/matconvnet/install/#compiling

2.run testSRnet_result.m for test result.

3.If you want to train it by yourself, you may download my data and use prepare_ur_data.m to produce imdb.mat which include every picture path.

4.Use train_SRnet.m to have fun~

(I also provide a verson for gray-scale images. But the improvement is limited. You can learn more from train_SRnet_gray.m and testSRnet_gray.m)

Small Tricks

1.I fix the scale factor 2(than 2+2*rand). It seems to be easy for net to learn more information.

2.How to initial net? (You can learn more from /matlab/+dagnn/@DagNN/initParam.m) In this work, the initial weight is important!

Citation

We greatly appreciate it if you can cite the website in your publications:

@misc{2016_super_resolution,
  title = {{2016_super_resolution}},
  howpublished = "\url{https://github.com/layumi/2016_super_resolution}",
}