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

have you met grid pattern in raw denoising training? #10

Closed dididadida closed 1 year ago

dididadida commented 1 year ago

could not fix this by tuning parameters in losses.

zejinwang commented 1 year ago

I can't understand what your problem is, please explain in detail.

dididadida commented 1 year ago

image image 第一个是输入,第二个是输出。基于我自己的数据训的。测试的时候,图片过完模型长这样,模型没收敛(训练越久越奇怪)。是得怎么微调那些训练参数呢?

zejinwang commented 1 year ago

raw域噪声数据训练时需要按照bayer阵列进行展开,由单通道变成四通道,请注意下这个细节

dididadida commented 1 year ago

是展开了的,但就是训的很奇怪。把你的模型和masker缝合到我的代码里训不出一个收敛的。但是我自己的代码训n2n没问题。我直接用你训练好的权重测试我的数据也没异常

zejinwang commented 1 year ago

可以关注下这三个点,分别是学习率设置;训练新数据集的梯度稳定性以及代码迁移可能存在问题(重点关注下masker操作)。你可以直接在我提供的源码上训练看下是否有这个问题