Super Resolution's the images by 3x using CNN
Sub-sampling the images by cropping the input image of 33X33 dimension and checking the correctnes by label of dimension 21X21.
Label of the image is generated by cropping the 21X21 regions from it sequentially.
After that upscaling the image by 3x by bicubic interpolation; followed by downscaling it by the same. And cropping the 33X33 from it.
We used 21X21 dimension because when we feed 33X33 images to our classifier it output 21X21 images.