Unfolding the Alternating Optimization for blind super resolution

Introduction

This repo is for OpenMMLab Algorithm Ecological Challenge, and paper is Unfolding the Alternating Optimization for blind super resolution

Training

We add some codes based on openmmlab/mmediting. You have two methods to train:

• pre-generate all training set by using our preprocess_div2k_dataset.py and use our dataset class bsr_folder_dataset.py
• generate low-quality images and kernels during training by defining the training pipeline in configuration file. We add a degradation class in augmentaion.py

You can just use two config files：

• dan_div2k_x4_gt_only.py. Remember replace augmentation.py with ours in mmediting
• DAN_DIV2K_x4_v2.py.

Remember to generate PCA encoder by running create_pca_encoder.py

Then, training command is the same as openmmlab/mmediting

Evaluation

The evaluation for two methods are the same because we define same pipelines. Just be careful with datasetloader we use.

Model weights: Baidu:eqp3