Training and test use different blur parameters

sovit-123 / image-deblurring-using-deep-learning

PyTorch implementation of image deblurring using deep learning. Use a simple convolutional autoencoder neural network to deblur Gaussian blurred images.

20 stars 5 forks source link

Training and test use different blur parameters #2

Open KMCzajkowski opened 1 year ago

KMCzajkowski commented 1 year ago

Hi, could you explain why add_gaussian_blur.py uses 51x51 kernel: https://github.com/sovit-123/image-deblurring-using-deep-learning/blob/master/src/add_gaussian_blur.py#L13 but testing (https://github.com/sovit-123/image-deblurring-using-deep-learning/blob/master/src/add_gaussian_blur_test_image.py#L18) uses 15x15 kernel? I tried training when they are both the same, but the results are bad in this case.

sovit-123 commented 1 year ago

@KMCzajkowski The add_gaussian_blur.py script was used to generate the training data. But the add_gaussian_blur_test_image.py is used to generate blurry images for inference. So, it can be any value of our choice. You can test on (51, 51) blurred images as well.