分层树结构字典编码的行为识别

发布时间：2018-01-10 20:26

本文关键词：分层树结构字典编码的行为识别　出处：《中国图象图形学报》2014年07期 　论文类型：期刊论文

【摘要】：目的基于学习字典的稀疏编码能够自适应地表示信号。然而,传统学习字典的原子之间缺少关联,信号的相似性在编码后缺失。考虑到结构化稀疏表示的鲁棒性和判别性能力,结构化字典的构建成为一个重要的任务。方法依据标准的凸优化字典学习算法,引入数据点编码路径的约束(由上层原子激活的索引规划下层的索引),构思了一种树结构字典学习框架。结果实验结果表明,局部描述符的稀疏表示具有较好的鲁棒性和判别性,同时在KTH数据库上人体行为识别实验与其他类似文献方法相比获得了较高的识别精度,其中,时空梯度方向直方图(HOG3D)的编码识别结果达到97.99%。结论通过实验结果,观察到采用本文构建的字典编码信号具有较好的鲁棒性和判别性,更好的适合分类任务。
[Abstract]:Objective to study the sparse encoding dictionary can adaptively based on signal. However, the lack of correlation between the traditional learning dictionary of atoms, the lack of similarity in signal encoding. After taking into account the structured sparse representation of robustness and discriminative ability, build a structured dictionary is an important task. The dictionary learning algorithm of convex optimization method according to the standard, the data encoding path constraints (lower index planning activated by the index of the atomic layer), the idea of a tree structured dictionary learning framework. The experimental results show that the sparse local descriptor representation has better robustness and discrimination, while in the KTH database and other human behavior recognition experiment similar methods in the literature compared to obtain a higher recognition accuracy, the spatio-temporal gradient direction histogram (HOG3D) encoding the identification results of the reach 97.99%. The test results show that the dictionary coding signal constructed in this paper has good robustness and discriminability, and it is better suited for classification tasks.

【作者单位】：东南大学信息科学与工程学院;
【基金】：国家自然科学基金项目(60971098,60672094)
【分类号】：TN911.7
【正文快照】： 0引言采用传统的方法对海量的图像或视频数据处理、分析、理解、存储和传输已不能满足当今需求,如何将这些庞大的数据实现简洁、自适应的表达,同时消除人类理解与底层数据之间的鸿沟成为人们日益关注的问题之一。近年来,海量数据的稀疏表示已有较好的发展,表现出了重要的研究

【相似文献】