基于改进的稀疏深度信念网络的人脸识别方法  

Face recognition algorithm based on improved sparse deep belief networks

作　　者：柴瑞敏[1] 曹振基 

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105 [2]辽宁工程技术大学研究生学院,辽宁葫芦岛125105

出　　处：《计算机应用研究》2015年第7期2179-2183,共5页Application Research of Computers

摘　　要：由于稀疏表示在人脸识别上的优异表现,大量的研究关注于在深度网络上结合稀疏编码。常用的稀疏深度信念网络限制所有的隐藏单元具有相同的稀疏水平,这不是诱导稀疏表示最自然的方式。针对这个问题,根据压缩感知理论改进原来的稀疏项,添加了一个tan-sigmoid正则项逼近稀疏表示的最优解L0范数。这种方法不限制隐藏单元拥有相同的激活率,每个隐藏单元可以根据不同的任务自动学习到不同的稀疏水平。在ORL、UMIST和FERET人脸库上的识别结果表明,提出的方法与经典的稀疏深度模型相比,获得了很好的特征表示和识别效果。Due to the excellent performance of sparse representation on the face recognition, a large number of studies focus on the deep network combining with sparse coding. Commonly used sparse deep belief networks limited all the hidden unit with the same level of sparse, which was not the most natural way to induce the sparse representation. To address this problem, according to the theory of compressive sensing to improve the original sparse items, this paper added a tan-sigmoid regnlarization, which was approaching to the sparse representation named the optimal solution of LO norm. This method did not limit the hidden unit with the same activation rate, and each hidden unit could automatically learn different levels of sparse according to the different tasks. The results in ORL, UMIST and FERET of human face recognition show that the proposed method, compared with the classical sparse depth model, obtains the very good feature representation and recognition effect.

关 键 词：稀疏编码 特征提取 深度学习 深度信念网络 稀疏受限玻尔兹曼机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]