search

JingyunLiang / SwinIR

SwinIR: Image Restoration Using Swin Transformer (official repository)
https://arxiv.org/abs/2108.10257
Apache License 2.0
4.44k stars 547 forks source link

JPEG CAR(Compression Artifacts Removal) for Colored Images??? #79

Open kewlcoder opened 2 years ago

kewlcoder commented 2 years ago

Does Swin_IR also work for colored JPEG images for JPEG_CAR problem? If yes, how to use it?

JingyunLiang commented 2 years ago

No. We haven't trained yet. We just followed the common setting in JPEG_CAR. React with thumbs up here if you want such a model. I will train and release one if there are more than 15 thumbs up. Or you can train one yourself with a few code modification.

rituational commented 2 years ago

Hi @JingyunLiang

It will be great if SwinIR can support CAR on color images. From the likes, it looks like many people are interested in the same. Let me know if you have plans to retrain the model to support this. It will be a great help. Thanks

JingyunLiang commented 2 years ago

Great! I will train it today. Do you want four separate models like 006_CAR_DFWB_s126w7_SwinIR-M_jpeg10.pth 006_CAR_DFWB_s126w7_SwinIR-M_jpeg20.pth 006_CAR_DFWB_s126w7_SwinIR-M_jpeg30.pth 006_CAR_DFWB_s126w7_SwinIR-M_jpeg40.pth

Or a joint model for JPEG10~40?

rituational commented 2 years ago

I think separate will be better if the quality of restored images is better that way. Do you think having separate models will help?

JingyunLiang commented 2 years ago

Yes, I agree. Training separate models have best performance. I will update here when the RGB models are released.

Siannodel commented 2 years ago

I'm also looking forward to it! Thanks!

rituational commented 2 years ago

Hi @JingyunLiang, any updates on this? Thanks

JingyunLiang commented 2 years ago

See results for Colored JPEG Compression Artifacts Removal in https://github.com/JingyunLiang/SwinIR#results and pretrained models at https://github.com/JingyunLiang/SwinIR/releases.

rituational commented 2 years ago

Hi, Thanks again @JingyunLiang for adding this. A minor thing. The testsets/LIVE folder doesn't exist. It is being called in the sample code you provided: python main_test_swinir.py --task color_jpeg_car --jpeg 10 --model_path model_zoo/swinir/006_colorCAR_DFWB_s126w7_SwinIR-M_jpeg10.pth --folder_gt testsets/LIVE1 but it gives error that no image found since the folder is empty.

rituational commented 2 years ago

Hi @JingyunLiang, you have provided small, medium and large variants for super resolution. For CAR, was there any specific reason for using the medium variant? (If I am not wrong, you have not provided small version right?) Thanks.