摘要
由于传统稀疏字典训练方法不能充分利用图像细节信息,提出一种分类稀疏字典训练方法。根据待训练样本的特性,将其划分为平滑、边缘和纹理三类,用KSVD算法分别训练出适合三类图像块特性的冗余字典,利用构造的冗余字典分别稀疏表示三类图像块。同时根据每类图像块所含信息量,自适应地分配测量率。实验结果表明,和单一正交基、冗余字典相比,该算法的稀疏系数更加稀疏,在低图像测量率时,重构效果更好,对边缘信息丰富的图像重构效果改善尤为明显。
Traditional learning dictionary can not reflect all detail information of origional image, so this paper proposes a new kind of learning dictionary, which is based on the classification of training samples. According to the characteristics of training samples, the proposed learning dictionary is able to divide the training samples into the smooth, the edge and the texture categories which will be trained by KSVD algorithm to get overcomplete dictionary. Block images can be classified into the smooth, the edge and the texture categories and sparsely represented by the proposed learning dictionary.Compared to single orthogonal base and traditional learning dictionary, the new proposed learning dictionary which can achieve more sparse representations is still able to give a good reconstruction performance at low sampling rates, and it can give higher image reconstruction performance for images with rich edge information.
出处
《计算机工程与应用》
CSCD
北大核心
2015年第6期193-198,共6页
Computer Engineering and Applications
基金
国家自然科学基金(No.61271240)
江苏省2013年度普通高校研究生科研创新计划项目(No.CXZZ13_0491)
关键词
分块压缩感知
自适应测量
分类稀疏表示
冗余字典
block compressed sensing
adaptive measure
classified sparse representation
overcomplete dictionary