This a unofficial testing only script written for ARCNN. It loads pretrained weights (.m) and execute them in PyTorch. There is absolutely no training functionality.
For details, please refer to the paper author's website: Deep Convolution Networks for Compression Artifacts Reduction.
This can act as a pre-processing step to other image processing algorithm like the ESRGAN which suffers from jpeg artifact.
python run.py --dir [Folder for input images] --output [Output directory] --batch_size [Batch size] --quality [Quality of JPEG images used to train: 10/20/30/40]
All of these parameters are optional. By default, it will read images in test_imgs/ and output to results/.
If you encountered out of memory, please lower the batch size. If the input images are of different resolutions, they cannot be batched and batch size 1 should be used.
GPU will be used if one is present.
Speedwise, tt can process a 224*224 image within 5ms with a GTX 1080Ti compared to around 3s/image in the matlab version.
In terms of quality, it has almost the exact same output as the author's Matlab testing code. You can use the script in the compare/ folder to verify this. Maximum absolute error is 1 and mean absolute error is 2e-3.
| Original | Author's matlab code | Our result |
|---|---|---|
 |
 |
 |
The network structure and pretrained weights are provided in the said project. The pretrained weights are uploaded here for convenience.