dcscn-super-resolution
dcscn-super-resolution copied to clipboard
Training Data Set
First off Thank you for sharing this work.
I have two questions, please?
-
Training data set do I need to convert it to YCbCr? or convert_y.py will take care of that.
-
How should I change for evaluation when I have my own LR image? by another way, I do need the code to generate low-resolution image.
Thank you
Hi, you don't need to convert to YCbCr for training. You can just put your RGB - HR image in the directory. That's same for evaluation. If you give HR images as an evaluation set, it will generate LR image from original HR image, and apply SR to create HR result image. Then it will compare the PSNR between original and SR result.
Does it answer your question?
Thank you for replying, my question was: I am trying to add specific noise to the images, is there any way to feed LR images ( which I create them ) and HR images for training? Because the code somehow creating down-sapling and noise to get LR.
Thank you
You mean, you want to add a noise when creating a LR image?
I use util.resize_image_by_pil() for both up scaling and downscaling. Probably you can modify it so that it add noise when the scale is below 1 (down sampling). Or you can also provide your own LR image for training. You can add a noise by your self in that case?
Because I used IM ( image model) some function in this model is P files ( I can’t see what is inside them but I know how these work) all made by MATLAB. So I am asking for way I can create my LR images in one file and HR in another file and let the code call them and use them for training. I hope this Explanation make my question little bit clearer.
Thank you
is it possible to make the scale = 1 ???