Golbstein / EDSR-Keras

EDSR Super Resolution in Keras

EDSR-Keras

EDSR Super-Resolution Implementation with Keras

Keras implementation of the paper "Enhanced Deep Residual Networks for Single Image Super-Resolution" from CVPRW 2017, 2nd NTIRE: EDSR Paper

Model Architecture

Extensions

1. Training with multi loss - MAE + VGG16 Perceptual Loss
2. float16 and float32 support
3. Keras Subpixel (Pixel-Shuffle layer) from: Keras-Subpixel
4. ICNR weights initialization - Checkerboard artifact free sub pixel convolution initialization, credit also for @kostyaev for the implementation of the initializer here: https://github.com/kostyaev/ICNR

Training from scratch in float16 with multi-loss doesn't work. Set to float32

Dependencies

• Python 3.6
• Keras>2.0.x
• keras-tqdm (pip install keras-tqdm)

Results of multi-loss

• Dataset: Pascal VOC 2012
• n_feats = 64
• n_resblocks = 8