| 原子稀疏结合块结构稀疏的联合表示图像识别算法 |
| |
| 引用本文： | 胡正平,宋淑芬.原子稀疏结合块结构稀疏的联合表示图像识别算法[J].信号处理,2013,29(7):888-895. |
| |
| 作者姓名： | 胡正平 宋淑芬 |
| |
| 作者单位： | 燕山大学信息科学与工程学院 |
| |
| 基金项目： | 国家自然科学基金(No.61071199);河北省自然科学基金(No.F2010001297);中国博士后自然科学基金(No.20080440124);第二批中国博士后基金(No.200902356)特别资助 |
| |
| 摘 要： | 针对结构稀疏表示识别算法中稀疏准则的选择以及字典内块的划分两个重要问题，提出两种改进的结构稀疏表示识别算法。首先，针对结构稀疏准则会出现较多系数不为零的情况，提出将结构稀疏准则与原子稀疏准则相结合的思路，包括并行和串行两种结合方式。并行结合是将两者以加权求和的方式同时作为稀疏表示的判别准则进行分类，串行结合则是在结构稀疏表示后，通过重组字典，再对测试样本进行原子稀疏表示实现分类。然后，针对字典中类内样本的块划分问题，提出基于MLP的结构稀疏表示识别算法，先将类内样本经过MLP的划分，保证各个分块分别位于低维的线性子空间中，再进行结构稀疏表示的分类。实验结果证明两种改进的结构稀疏表示识别算法的有效性。
|
| 关 键 词： | 图像识别 稀疏表示 结构稀疏 原子稀疏 最大线性块 |
| 收稿时间： | 2012-11-27 |
Sparse representation algorithm for image recognition based on the combination of structured sparse and atom sparse |
| |
| HU Zheng-ping,SONG Shu-fen.Sparse representation algorithm for image recognition based on the combination of structured sparse and atom sparse[J].Signal Processing,2013,29(7):888-895. |
| |
| Authors: | HU Zheng-ping SONG Shu-fen |
| |
| Affiliation: | School of Information Science and Engineering, Yanshan University, Qinhuangdao |
| |
| Abstract: | Two modified structured sparse representation algorithm for image recognition is proposed aiming at the two problems that the selection of the sparse criterion and the blocks’ division in the dictionary. First of all, according to there are more coefficients that is not zero in the structured sparse criterion, the thought of combining the structured sparse criterion and the atom sparse criterion is proposed, including both parallel and serial manner. In parallel combination, weighted summation of the both is used as the discriminant criterion. In serial combination, the dictionary is reconstructed after structured sparse representation, and then the atom sparse representation is used to achieve classification. Then, according to the problem that the blocks’ division of samples from the same class in the dictionary, structured sparse representation algorithm for recognition based on MLP is proposed. The images in the same class are divided into blocks based on MLP first to ensure that each block lies in low dimension linear subspaces respectively. After that, the test image is recognized by structured sparse representation. Experimental results show that both of the algorithms are effective. |
| |
| Keywords: | |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《信号处理》浏览原始摘要信息 |
|
点击此处可从《信号处理》下载全文 |
