search

tegg89 / SRCNN-Tensorflow

Image Super-Resolution Using Deep Convolutional Networks in Tensorflow https://arxiv.org/abs/1501.00092v3
MIT License
595 stars 279 forks source link

Different PSNR values from the paper #25

Open chensteven opened 6 years ago

chensteven commented 6 years ago

After training and testing it, I'm getting a different PSNR value which is lower than the one reported from the original paper. Is anyone able to replicate the same PSNR?

matt-dong commented 6 years ago

To get the same PSNR you would have to follow their steps very precisely, including the days of training types with expensive GPUs. It's expected to get a lower PSNR value since you haven't trained as long as they did.

niuniu111 commented 6 years ago

I didn't even see anything after the test. I didn't know what was going on.

fredlll commented 5 years ago

@chensteven how do you calculate the PSNR? Since the output images have different sizes as the orginal images.

zhaoyucong commented 5 years ago

@chensteven how do you calculate the PSNR?The PSNR that I came up with is two Numbers smaller than the average PSNR in the paper, that's a big difference can you share your code,

my email 1606435031@qq.com

15732031137 commented 4 years ago 
Hello! First of all, thank you for sharing your thoughts on the SRCNN program. I am from China and I am still a beginner. I recently reproduced SRCNN, but I encountered a problem. Because the output image after testing is not the same size as the original image, I cannot Calculate psnr, so I want to ask you how can I solve this problem? I wish you a happy life and academic progress! Looking forward to your review!
Newone48 commented 4 years ago

Sorry that I cannot help,because I am a new beginner too.You can ask somebody else,Good luck.

------------------ 原始邮件 ------------------ 发件人: "15732031137"<notifications@github.com>; 发送时间: 2020年3月1日(星期天) 上午10:51 收件人: "tegg89/SRCNN-Tensorflow"<SRCNN-Tensorflow@noreply.github.com>; 抄送: "Subscribed"<subscribed@noreply.github.com>; 主题: Re: [tegg89/SRCNN-Tensorflow] Different PSNR values from the paper (#25)

Hello! First of all, thank you for sharing your thoughts on the SRCNN program. I am from China and I am still a beginner. I recently reproduced SRCNN, but I encountered a problem. Because the output image after testing is not the same size as the original image, I cannot Calculate psnr, so I want to ask you how can I solve this problem? I wish you a happy life and academic progress! Looking forward to your review!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

YeHuanjie commented 4 years ago

did u solve this problem? the psnr of my bicubic image on SET5 is only 27.27db(for scale x4), while 28.42 in the paper. and that in my srcnn image is 28.28db while 30.48 in paper. how can i get 28.42db for bicubic image.

chensteven commented 4 years ago

Hey guys, @MashuPichu has the right idea. Your PSNR will generally be lower unless you follow the exact procedure and settings as the paper. Settings that will impact the PSNR include the number of back-propagation during training, dataset used for training and etc.

Second, the predicted image does come out smaller. One way to handle this is to crop the original image during the PSNR calculation.

I notice the original author of the code is no longer active. I'm currently working on a re-implementation of the SRCNN paper using Python3 and Tensorflow 2. I will also try to resolve these issues and document the solutions explicitly. I will keep you guys posted and release the code soon.

Edited: Suggest crop instead of pad for second point

jb2020-super commented 3 years ago

did u solve this problem? the psnr of my bicubic image on SET5 is only 27.27db(for scale x4), while 28.42 in the paper. and that in my srcnn image is 28.28db while 30.48 in paper. how can i get 28.42db for bicubic image.

You should use the imresize of MATLAB! Since I do not have MATLAB, I used Octave instead, and got an almost similar result for Set5: 2x: 33.8375 dB 3x: 30.0349 dB 4x: 27.7686 dB compared with the paper: 2x: 33.66 dB 3x: 30.39 dB 4x: 28.42 dB

zhangyunqaq commented 1 year ago

你解决了这个问题吗？我的双三次图像在 SET5 上的 psnr 仅为 27.27db（对于 scale x4），而论文中为 28.42。在我的 srcnn 图像中是 28.28db，而在纸上是 30.48。我怎样才能得到 28.42db 的双三次图像。

你应该使用 MATLAB 的imresize！ 由于我没有 MATLAB，所以我改用 Octave，得到了与 Set5 几乎相似的结果： 2x：33.8375 dB 3x：30.0349 dB 4x：27.7686 dB 与论文相比： 2x：33.66 dB 3x：30.39 dB 4x：28.42 dB

您好，可以请问一下您的PSNR是怎么计算的吗？万分感谢