Super_resolution icon indicating copy to clipboard operation
Super_resolution copied to clipboard

优酷阿里超分辨率重构比赛

RCAN

This repository is implementation of the "Image Super-Resolution Using Very Deep Residual Channel Attention Networks".

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.
In addition, we use a intermediate weights because training process need to take a looong time on my computer. 😭

Original
BICUBIC x2
RCAN x2

Usages

Train

When training begins, the model weights will be saved every epoch.
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