A Regularized Neural Optimization Approach with Beta Densities for Blind Deconvolution

In the field of image blind deblurring, existing deep learning deblurring networks can learn from a large number of paired training images, but are limited in dealing with a variety of complex large-size blur kernels. The underlying reason is that, those networks are not sufficiently generalized to unseen blurs, and hence, it is necessary to explore a proper regularizer so as to better constrain the problem. In this paper, we are inspired by the Beta probability density function, and empirically find that it has a great potential in discriminating between the blurred images and sharp images. With such an observation, we try to investigate the prior under the maximum-a-posterior framework so as to benefit the deep learning methods. In specific, this paper proposes a regularized neural optimization approach with beta densities for blind deblurring. In difference to most existing deep learning methods, the proposed approach has greatly benefited from the model-based scheme in terms of both hand-crafted and self-learned priors. With amounts of experiments on synthetic datasets as well as real datasets, and we find that the recovered images show good performance, the proposed method is shown comparative or better performance than state-of-art algorithms.