基于过完备字典学习的全变分图像去噪方法  

A total variation image denoising method over complete dictionary learning

作　　者：张晓严 马杰[1,2] 武利涛[1] 

机构地区：[1]河北工业大学电子信息工程学院,天津300401 [2]天津市电子材料与器件重点实验室,天津300401

出　　处：《河北工业大学学报》2017年第1期1-7,17,共8页Journal of Hebei University of Technology

基　　金：国家自然科学基金(61203245);河北省自然科学基金(F2012202027)

摘　　要：为了改善低信噪比情况下去噪效果、边缘保持能力差的问题,提出一种联合全变分正则项的字典学习图像去噪方法.首先,把增广拉格朗日乘子法和正交匹配追踪这两种求解稀疏编码的方法跟经典的K-SVD思想相交融,改善字典性能;其次,将全变分去噪模型融入到基于字典学习的图像去噪理论中,在图像重构基础上,引入全变分约束项,作为改进去噪模型中新的一项,达到对噪声和图像边缘作后续优化处理、改善图像去噪性能的目的.实验结果表明,改进的去噪方法,在保持原有去噪效果前提下,在噪声标准差较大或者图片边缘信息丰富时,去噪图像更加自然,边缘更加清晰,视觉效果较好.In order to improve the low SNR on denoising ability of edge preserving image, proposed K-SVD dictionary learning and total variation regularization denoising method. Firstly, the augmented Lagrange multiplier method and orthogonal method of the two kinds of solving sparse encoding tracking matching with K-SYD classic thoughts to improve the performance of the dictionary; the total variational denoising model into image dictionary learning denoising theory improved image denoising performance. The experimental results show that the improved denoising method, while maintaining the original denoising effect, the noise standard deviation is larger or for the rich image edge information, image denoising is more natural with clearer edge, and the visual effect is better.

关 键 词：字典学习 图像去噪 增广拉格朗日乘子法 正交匹配追踪 全变分 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]