owxiaom / calc_psnr

A multi-process script to calculate image PSNR, based on SWINIR's code implementation.

Calculation of Peak Signal-to-Noise Ratio and Structural Similarity between two pictures

The scrip is extracted from SWINIR And it should be applied on the saving format of EDSR

Opencv-Python, torch (cpu is enough) are needed.

python psnr_only.py --task classical_sr --scale 4 --folder_gt  {benchmark_path}/benchmark/{benchmark_name}/HR --folder_sr {result_save_path}/results-{benchmark_name}

And I found that when calculting the metrics, it applys on image one by one, which is extremely slow, so I add a multi-process edition on the origin script, so we have psnr_multi.py. By tring, I found when the largeur of process pool is 20, it is the fastest for 5 benchmark datasets of super-resolution.

python psnr_multi.py --task classical_sr --scale 4 --folder_gt  {benchmark_path}/benchmark/{benchmark_name}/HR --folder_sr {result_save_path}/results-{benchmark_name}

At last I think when running the script on different benchmark, I have to change the benchmark name everytime and I found it annoying, so I write another file that helps me execute this script, which is ssim_shell.py, it also writes logs in the file path of certain experiment.

Just

python ssim_shell.py --exp_name {the saved experiment name} --scale {scale, scale is important when cropping the border}

Normally it finishes calculating on 5 benchmarks within 30 secondes.