I am applying for the grayscale image by using your "2016_super_resolution".

[star4s]

Hello. nice to meet you. I am studying super_resolution by using neural network. I am applying for the grayscale image by using your "2016_super_resolution". When I use your mat-file, net-epoch-4.mat , in the SRnet-v1-gray-128 folder for grayscale, it was more blurring and no higher resolution. I think that the reason is about Parameter. What kind of Parameter can I adjust or control? Can you tell me what the problem is ? In your program, there is no testSRnet_result.m for grayscale. So, I changed some parts for grayscale. For example, in testSRnet_result.m, img = zeros(128,128,3,batch_size,'single'); => img = zeros(128,128,1, batch_size,'single');

%im_1 = rgb2gray(im_1); => im_1 = rgb2gray(im_1);

thank you for your attention.

[layumi]

Dear @star4s : First, you can test my trained model by img = repmat(img,1,1,3); %this turn the gray image to the rgb img. I am not sure my model can be directly apply to your question. But you can try it first!

Second, if your data is quite different with the image from imagenet, I suggest you to retrain the model. It seems that you have trained your own net. You need to change the input channel from 3 to 1 and output channel from 3 to 1 as well. (You mentioned The testSRnet_result,m is trainSRnet_result.m???)

Some tricks:

1. How many training data do you have? You can do some data augmentation like mirror.
2. If your data is not different from the imagenet, you can use my model as pretrained model to finetune.

If you have any question, do not hesitate to contact with me. You are welcomed.

Zhedong