You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Recently I trained my own model on this dataset, which is based on your code. The PSNR and SSIM index is a bit higher than your benchmark. However, when I opened the test image generated by your test.py, I found that there are checkerboard artifacts in all images. The generated SR image is 1024 x 1024 and each checkerboard is 128 x128.
I think this might be relevant to the "inference()" in the test.py, as the input image is cropped to a 128 x 128 patch with a stride of 64. But when I tested your pre-trained model, this did not happen. Besides, in some images, the checkerboard appears clearly while in other images not.
Do you have any idea about this phenomenon?
What's more, could you please tell me some information about your samples, like what kinds of cell are they?
Thank you!
Zikang Xu
The text was updated successfully, but these errors were encountered:
It's indeed a bit weird, none of the SR networks we've tested have this issue.
For the inference code, we cropped the images into small patches to avoid the memory issue for big networks, and we perform overlapping on the patches to avoid this kind of artifacts that usually happen along the border of the SR results from the patches.
One easy fix is to run the inference on the whole image at one shot if the memory fits. Otherwise, maybe you can try larger patch size (e.g. 256x256) with a larger stride during inference.
Thank you for your reply! Here remains another question.
I trained your proposed JDSR network on the FMD dataset (Floursecnce Microscopy Dataset). Because it only contains noisy and noise-free image pairs, I downsampled the noisy images using MATLAB imresize. However, the PSNR and SSIM performances are much worse than RDN and ESRGAN. I am confused that whether this method is only suitable for real SR image pairs?
Hi,
Recently I trained my own model on this dataset, which is based on your code. The PSNR and SSIM index is a bit higher than your benchmark. However, when I opened the test image generated by your test.py, I found that there are checkerboard artifacts in all images. The generated SR image is 1024 x 1024 and each checkerboard is 128 x128.
I think this might be relevant to the "inference()" in the test.py, as the input image is cropped to a 128 x 128 patch with a stride of 64. But when I tested your pre-trained model, this did not happen. Besides, in some images, the checkerboard appears clearly while in other images not.
Do you have any idea about this phenomenon?
What's more, could you please tell me some information about your samples, like what kinds of cell are they?
Thank you!
Zikang Xu
The text was updated successfully, but these errors were encountered: