Super Resolution using EDSR. Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR) model trained to convert a Low-Resolution image to a Super-Resolution image.
- We train an EDSR model with low-resolution image as input and the same high-resolution image as output. Generator and Discriminator part is not used here.
- After training EDSR like above, we fine tune the high resolution (HR) images by building a model using EDSR as generator and a discriminator and train it giving LR and HR images pair as input and taking SR images as output.
DIV2K: https://data.vision.ee.ethz.ch/cvl/DIV2K/
- Training a super resolution model is difficult without a very powerful GPU.
- About EDSR, WSDR and SR-GAN models.
- About Perceptual loss and Pixel loss.
- This project is highly inspired by: http://krasserm.github.io/2019/09/04/super-resolution/#model-training
- Do check his project for depp explanation: https://github.com/krasserm/super-resolution
Note: I tried training the model on my GPU and failed miserably. Then I tried training it on Colab's GPU, and failed again. At last I had to use pre-trained weights on my model.