优酷阿里超分辨率重构比赛参赛源码+项目说明.zip

# RCAN This repository is implementation of the "Image Super-Resolution Using Very Deep Residual Channel Attention Networks". <center><img src="./figs/fig2.png"></center> <center><img src="./figs/fig3.png"></center> <center><img src="./figs/fig4.png"></center> ## Requirements - PyTorch - Tensorflow - tqdm - Numpy - Pillow **Tensorflow** is required for quickly fetching image in training phase. ## Results For below results, we set the number of residual groups as 6, the number of RCAB as 12, the number of features as 64. <br /> In addition, we use a intermediate weights because training process need to take a looong time on my computer. 😭<br /> <table> <tr> <td><center>Original</center></td> <td><center>BICUBIC x2</center></td> <td><center>RCAN x2</center></td> </tr> <tr> <td> <center><img src="./data/monarch.bmp" height="300"></center> </td> <td> <center><img src="./data/monarch_x2_bicubic.png" height="300"></center> </td> <td> <center><img src="./data/monarch_x2_RCAN.png" height="300"></center> </td> </tr> </table> ## Usages ### Train When training begins, the model weights will be saved every epoch. <br /> If you want to train quickly, you should use **--use_fast_loader** option. ### Test Output results consist of restored images by the BICUBIC and the RCAN. ### 体会 比赛最后生成的图片像素有点大，资源有限，自己掏钱买了一台1080ti电脑，吃了大半年土，奈何一块1080ti还是太low。 对一块显卡的人来说DBPN，RCAN层数必须要减少一下，否则测试会内存溢出，也试了下cvpr2019的SAN，也必须阉割一下。后面由于实习等原因放弃了，能学以致，这次比赛真的是不错的体验，一起加油吧