Latent semantic concept regularized model for blind image deconvolution

Blind image deconvolution refers to the recovery of a sharp image when the degradation processing is unknown. Many existing methods have the problem that they are designed to exploit low level image descriptors (e.g. image pixels or image gradient) only, rather than high-level latent semantic concepts, thus there is no guarantee of human visual perception. To address this problem, in this paper, a latent semantic concept regularized (LSCR) method is proposed to reduce the blind deconvolution problem at a semantic level. The proposed method explores the relationship between different image descriptors and exploits sparse measure to favor sharp images over blurry images. And matrix factorization is introduced to learn the latent concepts from the image descriptors. Then, the image prior can be described and constrained by the learned latent semantic concepts of image descriptors using a much more effective convolution matrix. In this case, the blind deconvolution problem can be regularized and the sharp version of the blurry image can be recovered at a new latent semantic level. Furthermore, an iterative algorithm is exploited to derive optimal solution. The proposed model is evaluated on two different datasets, including simulation dataset and real dataset, and state-of-the-art performance is achieved compared with other methods.