Dc-huang / Dual-channel-NIR-denosing

Through the idea of contrastive learning, the neural network structure of two Dncnn channels is designed to remove the mixed structural noise of infrared images. Pre-training is required in the training phase, and contrastive training is fine-tuning, and the entire model is more dependent on pre-training results.
1 stars 0 forks source link

train #1

Open xiaowowwwww opened 1 year ago

xiaowowwwww commented 1 year ago

Excuse me, I would like to ask whether the paired picture in it is itself paired, or it is obtained by adding noise later? Is this not allowed to blind noise removal? I see that blind noise removal can be set in the previous program, which is a little unclear. Looking forward to your reply, thank you!

Dc-huang commented 1 year ago

I don’t know if it can solve your doubts. In fact, the training and testing data sets are noise-containing images obtained by simulating the original images,  so the input needs to be paired to compare clear pictures and noise pictures with psnr or some parameters. But if the model is trained, then the input can only use noise pictures, that is, the two inputs can be the same data set, we only need to pay attention to the output pictures . And test_shice is the real noise-containing images, that is, there is no original clear image for them.That is to say, in order to see the denoising results of test_shice, I use it as the input data set of clean images and noisy images.

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2023年6月29日(星期四) 下午5:20 收件人: @.>; 抄送: @.***>; 主题: [Dc-huang/Dual-channel-NIR-denosing] train (Issue #1)

不好意思,我想问一下里面的配对图片是自己配对的,还是后来加了噪音得到的?这样是不是不允许盲噪声去除?我看到在前面的程序中可以设置盲噪声去除,这有点不清楚。期待您的回复,谢谢!

— 直接回复此邮件,在GitHub上查看,或者取消订阅. @.***与>.

xiaowowwwww commented 1 year ago

Sorry, I still have some questions about this pre-training, does it mean that I have to train in another program about DnCNN? I'm really sorry, I'm a little confused. Look forward to your reply, thank you!