search

tensorlayer / SRGAN

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
https://github.com/tensorlayer/tensorlayerx
3.29k stars 810 forks source link

Why do images do poorly when the resolution is less than 100 #104

Open huangqianfirst opened 6 years ago

huangqianfirst commented 6 years ago

I am confuse why some test image greater than 300pix can works well, but when i test images which resolution less than 100pix or around 100pixel , the texture detail can't recover while the model it trained by 9696 to 384384 . like below , lr is input, 4x_gen is output. I kown it lr images may has been lost many detil, but why result is so awful and how can i impove upscale result of images resolution around 100pixel. sence6_4x_gen_400_748 sence6_lr_100_187 3_4x_gen_300_508 3_lr_75_127

wangyifenggg commented 5 years ago

I met the same problems. Maybe using a different training dataset can solve it. @huangqianfirst

huangqianfirst commented 5 years ago

@wangyifenggg . Yes . I trained with high resolution image and low resolution with JEPG noise images. This problem has been resolved.

JunbinWang commented 5 years ago

@wangyifenggg . Yes . I trained with high resolution image and low resolution with JEPG noise images. This problem has been resolved.

hi @wangyifenggg does the SRGAN model can handle the image compression artifacts by training it with jpeg LR dataset ?

wangyifenggg commented 5 years ago

@wangyifenggg . Yes . I trained with high resolution image and low resolution with JEPG noise images. This problem has been resolved.

hi @wangyifenggg does the SRGAN model can handle the image compression artifacts by training it with jpeg LR dataset ?

If you want to make your own training dataset, it's important to choose an appropriate way to make your LR dataset. For example, if you use nearest interpolation, the LR pictures will not smooth enough, which will damage the output of the SRGAN model. Maybe you can use area interpolation to produce LR pictures which are more smooth, and you can get better training result.