Some questions for the artifacts remove

[HuangBiaoVivo]

Excuse me, I'd like to know that if i train EDSR by using the datesets of the Track 2, the trained model can remove some artifacts as the competition required.

[sanghyun-son]

Hello.

Can I ask which kind of artifacts you want to remove?

If you prepare appropriate training dataset, you can do artifacts removal + super-resolution at the same time.

For example, if you consist the training set with [low-resolution jpeg images and high-resolution clean png images], the trained model will learn the jpeg artifact removal.

Please let me know if you want more detailed answer.

Thank you.

[HuangBiaoVivo]

Thanks for your response very much. I tried the datasets, which used the 1,400 photographs as HR images and the corresponding uploading images by WeiXin as LR images. But the results was not good as the corresponding artifacts still existed. I'd like to know if the problem relies on the datasets or the model.

[sanghyun-son]

Although I am not familiar with the term WeiXin, it seems that WeiXin is a kind of social network service (or chatting application?) that can upload the images.

Did you try the super-resolution and artifact removal simultaneously?

In that case, I wonder how did you prepare the dataset.

I think there are two cases.

1) Downsample the images first and upload them to the WeiXin. Then, download the uploaded images to prepare low-resolution samples.

2) Upload the images to the WeiXin first. Then, download the uploaded images and downsample them to prepare low-resolution samples.

Also, it will be helpful for me if you can upload the sample training HR-LR pairs.

Thank you!

[HuangBiaoVivo]

Thank you! Just as you think, WeiXin is a kind of chatting application. And when you upload a image by WeiXin, if the image is large enough(exceed around 1080), the image will be compressed. And I'd like to train a network that can complete super-resolution and compressed artifacts removal. I just take the photographs(4032 * 3024) by mobile phones as HR and the corresponding uploaded image(compressed by WeiXin) as the LR. But the results of the trained network is not well. Of Course, I can prepare the datasets for you if it is helpful for you.

[sanghyun-son]

I would like to see some sample images from your dataset.

Could you please upload a sample HR-LR pair for your experiment?

[splinter21]

@HuangBiaoVivo I think the author means that you should prepare the LR-GT image pair, and LR should be made by bicubic first -> weixin-jpeg, not weixin-jpeg first -> download it and bicubic.

[dubeys78]

i want artifact data for applying artifact removal methods ,so please suggest where i found these type of datasets