why the reconstruction image is too poor?

sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)

MIT License

2.43k stars 670 forks source link

why the reconstruction image is too poor? #34

Open stojakovic312636841 opened 6 years ago

stojakovic312636841 commented 6 years ago

excuse me! I recently train the EDSR_baseline_x2.pt model that used your pytorch code. the reconstruction image is too poor. So I want to request for you, thank you! I used cmd is that --> python3 main.py --data_test Demo --scale 2 --pre_train ../experiment/model/EDSR_baseline_x2.pt --test_only --save_results

and got the reconstruction image is too poor: img034_x2_sr

the input image is

sanghyun-son commented 6 years ago

Hello.

It seems that your input image is not downsampled by bicubic interpolation method. (I guess you used nearest-neighbor, since there are many aliasing artifacts.)

Since our model mainly handles bicubic-downsampled input images, we cannot guarantee the super-resolution result from your input. (It seems weird, I know. This situation is a common problem in recent deep super-resolution models.)

If you are trying to apply our method to that kind of images, I recommend you to train (or fine-tune) our model with nearest-neighbor downsampling method.

Thank you.

stojakovic312636841 commented 6 years ago

thank you very much. i have solved this problem when i used your suggest.

suke27 commented 6 years ago

@thstkdgus35 do you mean EDSR is just suit for bicubic downsampleed images, if we get a input images, we can not know whether it is downsample or just noise images

stojakovic312636841 commented 6 years ago

@suke27 个人认为EDSR的input和target之间是bicubicu的放大关系，因此EDSR学习的就是bicubic放大功能。在running的时候，input任何形式的图片都可以，input的图片通过EDSR之后，等效bicubic放大。

而我出现的问题在于，我将原始图片target，没有通过bicubic方式缩小成input。而后，将input通过EDSR后得到的output与target进行对比。此时的target != output 是合理的

suke27 commented 6 years ago

只有bicubic 方式缩写的input 放大才会有比较好的效果，是这个意思么，如果通过其他方式（压缩降质）之类的就会效果很差？

stojakovic312636841 commented 6 years ago

@suke27 你可以随便使用自己的原始图片（从网络上随便找）不通过bicubic缩小，直接进行EDSR，看一看效果。

suke27 commented 6 years ago

不是

splinter21 commented 6 years ago

@stojakovic312636841 所以你试过用预训练的EDSR模型继续用非双立方缩小的input-GT pairs，训练出针对不同缩小方式的模型吗，这种思路是否可行，还是说EDSR这种结构就只能训练双立方缩小的input-GT pairs

splinter21 commented 6 years ago

或者基于预训练的ESDR模型，喂各种缩小方法的input-GT pairs，可能双立方测试跑分会下降，但是可以或多或少学习到各种缩小方式的超分重建？

stojakovic312636841 commented 6 years ago

@splinter22 我想你应该阅读一下这篇简报《Ntire 2017 challenge on single image superresolution_ Methods and results-CVPRW-2017》，明确说明了如何比赛。EDSR总共训练了6个模型，3个模型参加bicubic downscaling。又有3个模型参加unknown downscaling http://www.vision.ee.ethz.ch/~timofter/publications/Timofte-CVPRW-2017.pdf