Image Deblurring Method Based On GSM FoE Model

With the popularization of portable digital products in our life, the problem of image degradation gains extensive attentions. As the most common phenomenon in image degradation, image deblurring is becoming one of the hot topics of intensive interest, which can be applied in aerospace industry, national defense, biomedicine, traffic safety and so on. Motion blur is the usual degradation situation in image degradations, which result from relative movement between imaging device and target object in the process of capturing images.Blind deblurring problem is the prior target in this paper. Firstly, the research background and actuality is simply introduced. Then several common image prior constraints and blind deblurring methods are briefly explained. After that, a new image restoration algorithm based on GSM (Scale Mixtures of Gaussians) FoE (Field of Experts) model and image gradient fidelity term is proposed. Besides, a method of kernel estimation based on l2 norm and continuity is proposed.Heavy-tailed distribution is the most important statistical characteristics of natural images. Most image restoration uses gradient prior with fixed parameter to restore images, and does not take the high order prior of the natural images into account. Aiming at the above problems, we propose a new image restoration algorithm based on GSM FoE model and image gradient fidelity term. Firstly, by training images from natural image library, GSM FoE model learns the filters and responding parameters which contain high order prior of the natural image. Secondly, these learned results are applied in guiding image restoration, an image deblurring model based on GSM FOE model and image gradient fidelity term is proposed, which can be solve effectively by IRLS algorithm.Experimental results prove that the proposed deblurring method can not only remove blur and noise easily, but also keep the sharp edges and suppress ringing artifacts. Moreover, our image restoration method performs well even for large blurring kernels.