结合图像块联合聚类加权和混合分类器的非对齐稀疏表示识别方法

项目名称： 结合图像块联合聚类加权和混合分类器的非对齐稀疏表示识别方法

项目编号： No.61501035

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 何珺

作者单位： 北京师范大学

项目金额： 15万元

中文摘要： 稀疏表示识别（SRC）方法在约束环境下取得了较好的效果，但对实际非约束环境的适用性并不理想，其性能与待识别图像和样本的对齐程度密切相关。为此，本课题提出了“结合图像块联合聚类加权和混合分类器的非对齐稀疏表示识别方法”。首先将样本图像分割成若干图像块，利用动态联合稀疏表示对样本图像块进行统计分析、聚类和权重计算，再用局部K-SVD方法训练并构建图像块字典；同时利用稀疏表示误差和稀疏度评价待识别图像块个体和样本图像整体之间的关系，并综合表示误差和最大似然估计，设计混合分类器；最后针对非对齐问题，利用快速定位算法，设计基于图像块的联合稀疏表示方法的应用框架。本课题分析图像块之间的关系、图像块和目标整体之间的关系，提出基于图像块SRC的思路，并融合聚类、统计分析和快速定位算法，解决传统SRC方法应用中的对齐问题；从而，提高SRC方法的应用鲁棒性，并为SRC方法的研究提供参考。

中文关键词： 图像识别；稀疏表示识别方法；动态联合稀疏表示；聚类加权；混合分类器

英文摘要： Sparse representation based classification (SRC) method has achieved good results under constraint environment, but still cannot operate well for actual unconstrained condition, that may be much related to the alignment between the probe image and the samples. Aiming at that problem, this project proposes an unaligned “image-block SRC method” based on dynamic joint sparse –representation clustering weighted and hybrid classifier. Firstly, many image-blocks, segmented from a sample image, will be selected, clustered and weighted in accordance with dynamic joint sparse representation. Based on the clustered sets, a dictionary will be available with local K-SVD method. Then, the relationship between the individual tested image-block and the overall sample image will be evaluated with sparse representation error and sparsity, with that a hybrid classifier can be designed based on maximum likelihood estimation and minimum mean square error estimation. Finally, considering the problem of alignment, a framework based on the proposed image-block joint SRC method will be designed with the rapid positioning algorithm. In this project, firstly analyzing the relationship between sample image blocks, and relationship between individual tested block and overall target-images, we give a novel idea of image-block based SRC method by taking the advantage of the methods of clustering, statistical analysis and rapid positioning. Thus, with the proposed method and framework, we may solve the problem of alignment of the original SRC method and improve its robustness in application. What’s more, this research may provide reference for the further study on SRC method.

英文关键词： image recognition;sparse representation based classification method;dynamic joint sparse representation;cluster-weight;hybrid classifier