search

zejinwang / Blind2Unblind

This is an official implementation of the CVPR2022 paper "Blind2Unblind: Self-Supervised Image Denoising with Visible Blind Spots".
https://arxiv.org/abs/2203.06967
118 stars 13 forks source link

The loss increases with training, please help！ #19

Closed lvxuzijian closed 7 months ago

lvxuzijian commented 1 year ago

损失随着训练而增加，最终的图像去噪效果显示规则的条纹形状。

截屏2023-07-14 21 02 38
zejinwang commented 1 year ago

"I suspect that the noise in your data may have strong spatial correlation, and this method is primarily targeted at pixel-independent noise. For single-channel grayscale images, adding TV loss can effectively solve the problem you are facing."

lvxuzijian commented 1 year ago

Thanks for you reply! I will give it a try!

yashkhurana24 commented 7 months ago

@lvxuzijian what denoising method worked for you?