Residual-Channel-Attention-Network

Unofficial Tensorflow implementation of "Image Super-Resolution Using Very Deep Residual Channel Attention Networks", ECCV 2018, [arXiv]

Train

1. Training Dataset is 800 training + 100 validtion images of DIV2K dataset.The TFRecords for the same and npz files of the benchmark dataset can be found in the drive link.

2. Maintain the following Directory structure

3. cd to code. Specify the scale of upsampling by '--scale'. Default value is 4.

4. Run python3 training.py

Inference

Run python3 infer.py . The output will be stored in results.

Evaluation (NOTE : THIS REQUIRES MATLAB)

1. cd to eval

2. Run ./matlab_eval.sh

Sample Results for x4 upsampling

Note that all the results shown below are obtained from limited period of training. Accuracy can be improved when trained for more duration.

Ground Truth	Predicted

Results obtained are in Y channel. We can perform Bicubic interpolation in Cr and Cb channels and merge with the predicted Y channel.