mivrc / mlefgn-pytorch Goto Github PK

Hello ,my research interest is X-ray imaging and its applications.I want to use your code CT projection denoising because the excellent results in your paper.But what confusing me is that the data value fluctuate around 1.just like 1.000058 or 0.999954.
At the first,i tried tranforms my data to .png,it does works，resuls is pretty good.But the value is between 0-255,it can't reconstructed successfully because it losses many information.And I tried modify the code to input a matrix output a matrix with my true data value without normalized.But the L1 loss is too small and soon to zero.

Can you tell me how can I solve this issue?Like use normalization before the training or modify the loss function.Or if you are intrested in my area we can incooperate to do more thing.Like we can provide simulation and experimental dataset include attenution-based \phase contrast CT and expert knowledge. you can give us some advice about machine learning.
Best regards.

Use the following command:
python main.py --template MLEFGN --save MLEFGN_Gray_15 --scale 1 --reset --save_results --patch_size 64 --ext sep_reset
I got model_best.pt from Train/experiment/MLEFGN_Gray_15/model
But,when I use the following command to test,
python main.py --data_test MyImage --scale 1 --model MLEGN --pre_train ../model/model_best.pt --test_only --save_results --chop --save "MLEFGN_Gray_15" --testpath ../LR/LRBI --testset Set12
I got an edge detection image.

How can I get a image denoising model？

I see you have many loss function can be chosen,like adversarial loss and vgg loss.

And you also said it's not a GAN-based model,does it means if I use the adversarial loss,the original model MLEFGN can be seen as the 'generator',it generates the image to the discrimator.

And does the vgg loss means the perceptual loss ? But i didn't seen any pretrained model like vgg-19.

Thanks !

i've been successfully training in your default setting,but when i want to test,i can't find the .m doc.

And interestingly,i find the WGAN in the loss function,but i don't see it in u network,does it means even without the generator one can still use the adversial loss?

thank u very much.wonderful work

Hello,I read your paper throughly in a couple days, and run your code successfully,the result is amazing,but what confuse me is that the design of your Edge-Net.It seems like a common feature extraction or mapping network.

(Sorry I am new to this area,maybe have some inaccuracy description )

I wonder know how you eliminate low-frequency features and image noise simultaneously and remain the edge information.As I know the noise and edge information both are high-frequency features.

I made an environment with pytorch 0.4.1 and python 3.6 on ubuntu 18.
But i cannot run test code, I followed your instruction but i met the following message:
Exception has occurred: AttributeError
module 'model' has no attribute 'Model'
File "/.../MLEFGN-PyTorch-master/TestCode/code/main.py", line 15, in
model = model.Model(args, checkpoint)

What is the reason?
Please help me.

Hi, I'm trying to reproduce the testing result, but currently all I get is bizarre images as I show below. I'm using pre-trained model and testing data that already inside the repo.
My Environment:
Ubuntu 18.04
python 3.6.13
pytorch 0.4.0
torchvision 0.2.0
cuda 9.0

Command I run:

python main.py --data_test MyImage --scale 1 --model MLEFGN --pre_train ../model/15_gray.pt --test_only --save_results --chop --save "MLEFGN_Gray_15" --testpath ../LR/LRBI --testset Set12 --cpu

Input image:

Output images(denoised):

When I was tring to train the model with gray-scaled images, I put the gray-scaled images and noisy images into folders of "DIV2K_train_HR" and "DIV2K_train_LR_bicubic/x1". The code seems running, however, the loss just keeps increasing during the training, and becomes "nan" finally after about 50 epochs of training. Maybe I have made something wrong about the training data? Can you provide some suggestions about this? Thanks.