Systems and methods for training a non-blind blur removal module

摘要

Methods and systems for training a non-blind blur removal module are disclosed. Defocused test images and defocused test images are received, each of the defocused test images being related to a corresponding one of the defocused test images by a defocused kernel term and a noise term. A controlled deconvolution sub-module and a convolutional neural network are trained together by adjusting a control parameter of a controlled deconvolution function and weights of a convolutional neural network to minimize a cost function that represents a difference between each defocused output image and a corresponding one of the defocused test images.