多方向加权TV和变换域非局部正则化图像去模糊
Image Deblurring Based on Multi-Direction Weighted TV and Transform Domain Non-Local Regularization

针对传统全变差(TV)去模糊对噪声敏感且细节恢复能力有限等缺点, 利用边缘检测对传统TV模型进行改进, 并受空域非局部自相似性正则化思想启发, 将图像的变换域非局部自相似性约束融入去模糊模型, 提出一种基于边缘检测的多方向加权TV和变换域非局部正则化的图像去模糊方法．首先, 运用边缘检测将中心像素邻域内的像素对划分为同侧像素对和异侧像素对, 对不同类型的像素对采用不同的权重, 在去模糊的同时尽可能保持图像边缘等细节特征; 其次, 为充分利用先验信息, 将变换域非局部正则化约束融入到改进的TV模型, 进一步改善图像视觉质量; 最后, 对新模型进行有效求解．实验结果表明, 本文算法在去模糊的同时可更好地保留图像的边缘、纹理等细节特征．
In view of the shortcomings of noise sensitivity and limited recovery ability of traditional total variation (TV) model，an improved TV model is proposed. At the same time, inspired by the idea of spatial non-local self-similarity，the non-local regularization constraints of transform domain is integrated into the above model. So，the new image deblurring method is proposed based on mulit-directional weighted TV and regularization in transform domain. First of all，all pairs of pixels were divided into those which are on the same sides and those which are on opposite sides of an edge based on edge detection，then different weights were defined on different pairs of pixels. This method can deblur an image and preserve the image edge well. Furthermore，in order to use the prior infor-mation of the deblurred image，transform domain self-similarity regularization was introduced to the TV model to better preserve the details and texture. Then a modified alternating direction method was addressed to solve the above model. Experiments show that the proposed algorithm not only improves the visual quality，but also preserves details and texture well