Strange Unet artifact

1

I am using a Unet model (a simple encoder with average pooling and a decoder using ConvTranspose2d function) for image upsampling (super-resolution). A pair of images (input on the left and target on the middle) is used for training as shown below. Every other column from the target image is removed and zero-padded to come up with the input image.

Screenshot from 2024-10-11 23-34-14
Screenshot from 2024-10-11 23-34-141920×962 104 KB

During training, I could see the vertical line artifacts (image on the right) on the reconstructed image (model output) upon zooming in.

What can be done to rectify the artifacts? I am using L1 and L2 norms as loss functions. The training set contains 6000, 672 x 1024 grayscale images.

2

Your output and target look close to the same, just flipped vertically. You sure you aren’t flipping one of them before viewing?

3

Pretty big images, are you patching them (if so experiment a bit more here?) or feeding them whole?

4

@J_Johnson Yes, the output looks close to the target. I have uploaded the full-size validation target and output images at Unet — ImgBB. You can see the artifacts when you zoom in on the output.

@Soumya_Kundu, I am feeding the whole image. I will give reducing/patching the images a try and see if that works.

Thanks for the response guys!