원문정보
초록
영어
Up to now, the non-convex ℓp (0 < p < 1) norm regularization function has shown good performance for sparse signal processing. Indeed, it benefits from a significantly heavier-tailed hyper-Laplacian model, which is desirable in the context of image gradient distributions. Both ℓ1/2 and ℓ2/3 regularization methods have been given analytic solutions and fast closed-form thresholding formulae in recent image deconvolution methods. However, the methods with the other p-value norm penalty term still suffer difficulties in getting the analytic solution and fast closed-form thresholding algorithm. In this paper, to deal with these issues, we propose an approximation of ℓp regularization terms with 0.5 p < 1 using a linear combination of two ℓp terms (that is 1 and 1 / 2 ) with closed form thresholding formulae. We develop an alternating minimization method to solve the image deconvolution problems involving the constructed approximating function. We derive theoretical analytic solutions and fast closed-form thresholding formulae. 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, with an application to image deconvolution.
목차
1. Introduction
2. Non-convex ℓp Regularization Image Deconvolution
2.1. x Sub-problem
2.2. w Sub-problem
3. Our Algorithm
3.1. Approximation of the ℓp Regularization Term
3.2. Determining the Weight Coefficients of the Approximation of ℓp Norm
3.3. Solution to the Algorithm
4. Experiments and Analysis
5. Conclusion
Acknowledgments
References