摘要
提出了一种针对一类图像进行稀疏表示的字典训练方法,并证明了该算法的收敛性.该算法的几何解释是,以最少的超平面来逼近样本所在的一小块球冠.算法流程为聚类每一步迭代所产生的余项,将聚类中心作为新的字典原子,令字典能够更适应于样本的稀疏表示.该算法与传统的字典训练方法相比具有适应性强,对训练样本规模和字典规模要求低,收敛速度快,算法复杂度低等特点.利用该算法训练得到的字典用于压缩感知、图像去噪等实验表明,该字典具有很好的效果.
A dictionary training algorithm was proposed for spare representation of images and its convergence was proved.The geometrical explanation of the algorithm is to approximate the hyperspherical cap with least hyperplanes.The algorithm clustered the error vectors of each step,and signed the cluster center as new atoms which made the dictionary more suitable for spare representation of samples.Compared with the traditional algorithm,the new one has higher adaptability,lower requirement of sample number and dictionary size,higher convergence rate,and lower complexity.Finally,the experiment of compressive sensing and denoising demonstrates that dictionary training by this algorithm has good effect.
出处
《光子学报》
EI
CAS
CSCD
北大核心
2011年第2期316-320,共5页
Acta Photonica Sinica
基金
中央高校基本科研业务费专项资金(No.CHD2009JC156)
西安邮电学院青年教师基金(No.ZL2010-21)
长安大学基础研究支持计划专项基金资助
关键词
稀疏表示
聚类
压缩感知
字典
原子
稀疏度
Sparse representation
Clustering
compressive sensing
Dictionary
Atom
Sparseness