search

thangvubk / FEQE

Official code (Tensorflow) for paper "Fast and Efficient Image Quality Enhancement via Desubpixel Convolutional Neural Networks"
MIT License
128 stars 19 forks source link

About super resolution #10

Closed zhanzheng8585 closed 4 years ago

zhanzheng8585 commented 4 years ago

For super resolution, after downsampling, the feature maps shrink. And your code shows the reverse operation in upsampling stage. It seems that the final picture has the same size with original picture? I try to implement your code in pytorch, but the picture size didn't change.

for example

when scale = 2 downsample /2 upsample * 2

when scale = 4 downsample /2 downsample /2 upsample 2 upsample 2

thangvubk commented 4 years ago

Hi. This follow the definition of SRCNN for the definition of low-resolution image (which is bicubically interpolated images). You can see detail in section 3.1 of SRCNN paper.

zhanzheng8585 commented 4 years ago

Hi. This follow the definition of SRCNN for the definition of low-resolution image (which is bicubically interpolated images). You can see detail in section 3.1 of SRCNN paper.

Thank you!