Slight color(chroma) brightness change

[Lee-lithium]

Hello @jiaxi-jiang, Sorry to bother you,

I test some non-photo content jpeg(2d drawing),
high FBCNN QF can preserve detail and noise,
but I notice some area have slight color(chroma) brightness change.

In this case, FBCNN QF 70 have slight brightness change on dark red area(red circle),
could you teach me how to improve color accurate for non-photo content?

original image (jpeg q75 420),

png 8 bit depth (FBCNN QF 70)

Other sample(QF 30,50,70)

sample.zip

[jiaxi-jiang]

Hello, thanks for the proposed issue! From my observation, the brightness change is hard to distinguish, maybe just because the restored image is cleaner so looks different. On the other hand, blind jpeg image restoration is inherently an inverse problem, which means that for the same content in a compressed image, it can come from different uncompressed image by setting different quality factors. To eliminate this ambiguity, we learn prior knowledge from training data, but there must be difference of data distribution between training and testing images. So if you want to have better results on anime images, the best way is to directly train or finetune the model with anime images (I only use DIV2K and Flickr2K as training data). Besides, anime images usually don't contain much high-frequency information, so it should be enough to use my pre-trained model with very small qf to remove artifacts.

[Lee-lithium]

Thank you for your reply :)

maybe just because the restored image is cleaner so looks different.

That make sense, I think jpeg chroma subsampling(420) and jpeg q75,
create a lot of tiny chroma artifacts in this dark red area,
and jpeg ycbcr also have some issue for red color.

In my test, I think FBCNN QF 80~90 can let that dark red area restore better,
and current pre-trained model is really good for different content,

I plan implement input jpeg decode to uint16 and convert to float32 tensor,
result float32 convert uint16, and output png 16 bit depth,
maybe use high precision and reduce precision loss can let result better.

I think use libjxl Butteraugli 3-norm finetune model,
probably can improve FBCNN chroma performance.