Natural Image Deblurring Algorithms Based On Non-convex Regularization And Hardware Implementation

Image deblurring is a kind of technique,which can process the non-idea image affected by the outsider interference in order to get the image that has more details and is much closer to the original image. The image deblurring methods based on regularization mainly utilize the prior knowledge of the image to acquire more effective deblurring algorithms. Among them, The natural image deblurring methods based on regularization try to use the statistical features of natural images for image deblurring and get high quality restored image. This paper focuses on the research about natural image deblurring algorithms based on regularization and covers such content as natural image deblurring algorithms based on non-convex regularization and the specific implementation of algorithm in hardware system.In the respect of natural image deblurring algorithms based on non-convex regularization: Aimed at the selection of regularizers in the variational framework, We try to apply non-convex regularizers in convex total bounded variation models for image deblurring in this paper, and utilize a iterative reweighted method based on the Alternating Direction Method of Multipliers(ADMM) which is similar to the Iterative Weighted Least Squares Method(IRLS) to solve them. Systematic experimental results demonstrate that the proposed non-convex models which are applied to image deblurring show better performance and spend less CPU time; Besides, in view of the speed of the IRLS algorithm is too slow and the existing natural image deblurring algorithms(analytic solutions and closed-form thresholding formulae) based on non-convex lp regularization can only solve specific p values(p = 1/2; p = 2/3). We derive theoretical analytic solutions and fast closed-form thresholding formulae for the non-convex l0.8 regularization in this paper. Experiment results show that the proposed algorithm takes less computational time than IRLS method, while it has received a significant boost in computational efficiency. Further, We derive theoretical analytic solutions and fast closed-form thresholding formulae for a more general non-convex lp(0.5 ?p ?1) regularization in this paper. We perform extensive numerical experiments to demonstrate the versatility and effectiveness of the proposed method, through a comparison with the recent non-convex lp regularization dealing with the special p-value term(p = 1/2; p = 2/3).In the respect of hardware implementation: We realize the natural image deblurring core algorithms based on non-convex regularization on the TMS320C674 X DSP, hardware debugging is carried out by the emulator on the environment of CCS. Through analysing and comparing the results, we can find that the effects of processing on DSP board and on Matlab are basically identical, it can improve the processing speed of the algorithm significantly by using the DSP hardware acceleration.