Multi-Scale Low-Rank Blind Image Deblurring Method

Most existing single image blind deblurring methods based on statistical priors suffer a poor texture restoration and ringing artifacts. A multi-scale blind image deblurring method based on patch-wise local maximum gradient prior and low rank prior is proposed in this paper. Specifically, we choose a coarse-to-fine multi-scale framework to construct an image pyramid via gray-scale and down-sampling operations layer by layer. At the single-scale level, the patch-wise local maximum gradient prior and the low-rank prior are put into the MAP framework, and then both the intermediate latent image and blur kernel are estimated with the alternating direction method of multiplier and the half-quadratic splitting method. We finally obtain the sharp image by performing non-blind deconvolution for the blurred image and estimated kernel based on hyper-Laplacian prior and total variation-L2 method. Solving the low-rank regularization directly is computationally expensive, we thus transform the sub-problem of low-rank regularization term constrained by weighted Schatte-1/2 norm into a sub-problem of non-convex weighted L1/2-norm, and then adopt the generalized soft-thresholding method (GST) to achieve the global optimal solution. Experimental results and comparisons with the existing classical image deblurring methods on the benchmark datasets show that the proposed method facilitates a better image deblurring performance. After image deblurring on Köhler synthetic dataset, the average peak signal-to-noise ratio reaches 30.06 dB, the average structural similarity reaches 0.946 5, and the estimated blur kernel gets more accurate. © 2021, Editorial Office of Journal of Xi'an Jiaotong University. All right reserved.