电子学报 ›› 2009, Vol. 37 ›› Issue (2): 347-350.

• 论文 • 上一篇    下一篇

一种基于超完备字典学习的图像去噪方法

蔡泽民1,3, 赖剑煌2,3   

  1. 1. 中山大学数学与计算科学学院,广东广州 510275;2. 中山大学信息科学与技术学院,广东广州 510275;3. 广东省信息安全技术重点实验室,广东广州 510275
  • 收稿日期:2008-03-03 修回日期:2008-09-05 出版日期:2009-02-25 发布日期:2009-02-25

An Over-complete Learned Dictionary-Based Image De-noising Method

CAI Ze-min1,3, LAI Jian-huang2,3   

  1. 1. School of Mathematics and Computational Science,Sun Yat-sen University,Guangzhou,Guangdong 510275,China;2. School of Information Science and Technology,Sun Yat-sen University,Guangzhou,Guangdong 510275,China;3. Guangdong Province Key Laboratory of Information Security,Guangzhou,Guangdong 510275,China
  • Received:2008-03-03 Revised:2008-09-05 Online:2009-02-25 Published:2009-02-25

摘要:

基于超完备字典的图像稀疏表示因其具有稀疏性、特征保持性、可分性等特点而被广泛应用于图像处理.本文提出一种超完备字典学习算法并应用于图像去噪.将字典学习等价于一个二次规划问题,并提出适合于大规模运算的投影梯度算法.学习所得字典能有效描述图像特征.基于此超完备学习字典,获得图像的稀疏表示,并恢复原始图像.实验结果表明,与小波类去噪方法相比,本文的学习算法能更好地去除图像噪声,保留图像细节信息,获得更高的PSNR值.

关键词: 稀疏表示, 基追踪, 匹配追踪, 字典学习, 二次规划

Abstract:

Images’ sparse representations over over-complete dictionaries have a wide application in image processing due to the properties of sparsity,integrity and separability.This paper proposes a dictionary learning algorithm which is applied to image de-noising.The dictionary learning problem is expressed as a box-constrained quadratic program and a fast projected gradient method is introduced to solve it.The learned dictionary describes the image content effectively.Experimental results show that:in comparison with the wavelet-based de-noising methods,our learning-based algorithm has better de-noising ability,keep more detail image information and improve the peak signal-to-noise ratio.

Key words: sparse representation, basis pursuit, matching pursuit, dictionary learning, quadratic program

中图分类号: