Closed huizhang0110 closed 5 years ago
The blur kernels were created using Matlab's fspecial
(https://uk.mathworks.com/help/images/ref/fspecial.html). Several lengths and angles were used in order to capture different blurs.
In this case, will the network be over-fitting to such predefined 2-D filter?
Yes, it will eventually, just as every network would eventually overfit the training data given a large enough network size and long enough training. Deblurring is all about artificially modelling realistic blurs as closely as possible, so if you knew the blur kernel that occurred in reality, you'd want to overfit on this kernel. However, this kernel is unknown, so my approach was to train on a large variety of blur kernels. This prevents overfitting to one specific motion blur kernel. However, it does overfit to linear motion blur kernels (as I have only used linear motion blur), so the trained network could not be used for deblurring images that have been blurred with other kernels.
Thanks for your code, but are these blur kernels randomly generated?