search

caiyuanhao1998 / Retinexformer

"Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement" (ICCV 2023) & (NTIRE 2024 Challenge)
https://arxiv.org/abs/2303.06705
MIT License
920 stars 81 forks source link

Testing on images with lots of noise returns bad result #121

Closed willswordh closed 1 month ago

willswordh commented 1 month ago

@caiyuanhao1998 Hi Yuanhao, I am testing to use your model to process low-light pictures that have lots of noise which are shot from not-advanced cameras. The direct result is not good. Do you recommend to apply some deep-learning based denoise model first to the image, then use your model to process it? Or will it make the result even worse? Hope to get some insight from you, thanks a lot!

caiyuanhao1998 commented 1 month ago

Hi, it would be better you switch to our other enhancement models that can handle severe image degradation

I suggest you try SID and SMID models.

If you found our repo useful, please help us star it, thanks :)

willswordh commented 1 month ago

@caiyuanhao1998 Hi Yuanhao, SID and SMID models are for raw pictures. Is my understanding right that for pictures like jpg, the general low light enhancement models trained on LOL like datasets will not be able to handle severe image degradation? I am thinking to apply some denoise models based on jpg pictures to eliminate noise then apply your model to do enhancement. Not sure if it is gonna work out. Please share your insight. Thanks!

willswordh commented 3 weeks ago

@caiyuanhao1998 Please offer your valuable insight if you are able to see this