An implementation of SRCNN using pytorch The orignal paper is Image Super-Resolution Using Deep Convolutional Networks here
- Adam is used instead of SGD
- Removed weight initialization
- Trained on a different dataset. The dataset is available here: BaiduYun
The result is better than bicubic to enlarge a 2x pic.
Follow are the bicubic-img and srcnn-img
psnr of Bicuibic image = 26.492; psnr of SRCNN image = 28.687
python train.py --output-dir './result' --optimizer Adam --num-epoch 50 --batch-size 128
python test.py --image-file test/4.jpg --net-file result/Adam/final_net.pth