基于重叠稀疏组深度信念网络的图像识别

计算机工程与科学

基于重叠稀疏组深度信念网络的图像识别

田进，陈秀宏，傅俊鹏，徐德荣

（江南大学数字媒体学院，江苏无锡 214122）

收稿日期:2016-04-19 修回日期:2016-09-30 出版日期:2018-03-25 发布日期:2018-03-25
基金资助:
国家自然科学基金（61373055）；江苏省2015年度普通高校研究生实践创新项目（SJLX15-0566）

Image recognition based on an overlap

sparse group deep belief network

TIAN Jin,CHEN Xiuhong,FU Junpeng,XU Derong

(School of Digital Media,Jiangnan University,Wuxi 214122,China)

Received:2016-04-19 Revised:2016-09-30 Online:2018-03-25 Published:2018-03-25

摘要/Abstract

摘要：

深度信念网络的隐含神经元大部分为噪声变量，且具有组结构相关性。组稀疏深度信念网络模型通过组Lasso模型对隐含神经元变量进行约束，从而实现变量组选择。然而，组稀疏深度信念网络模型未能考虑特征可同时属于多个特征组，并且隐含神经元在变量层面上不稀疏的问题。在组稀疏深度信念网络模型上引入重叠组结构，解释了重叠组Lasso模型在变量层面上比组Lasso模型稀疏的原因，并在变量层面上作进一步的稀疏，提出了重叠稀疏组深度信念网络模型。在MNIST、USPS、ETH-80以及人脸数据集上的识别结果表明，重叠稀疏组深度信念网络具有更高的识别率。

关键词: 深度信念网络, 组Lasso, 组稀疏, 重叠稀疏组

Abstract:

Most hidden neurons in Deep Belief Network (DBN) are noise variables, and have group structure correlation. Group Sparse Deep Belief Network (GSDBN) constrains the implicit neuron variables via group Lasso so as to achieve variable group selection. However, the group sparse model not only ignores the case that some features belong to multiple groups simultaneously, but also has the problem that hidden neurons are not sparse. In this paper, we propose Overlap Sparse Group Deep Belief Network (OSGDBN), which introduces the overlap group structure and makes the hidden neurons sparse, based on Group Sparse Deep Belief Network. In addition, we also explain the reason that the OSGDBN is sparser than GSDBN. The recognition results on MNIST, USPS, ETH80 and face datasets show that OSGDBN has a higher recognition rate.

Key words: deep belief network, group Lasso, group sparse, overlap sparse group

田进，陈秀宏，傅俊鹏，徐德荣. 基于重叠稀疏组深度信念网络的图像识别[J]. 计算机工程与科学.

TIAN Jin,CHEN Xiuhong,FU Junpeng,XU Derong.

Image recognition based on an overlap

sparse group deep belief network

[J]. Computer Engineering & Science.

[1]	肖汉雄，陈秀宏，田进. 特征聚类自适应变组稀疏自编码网络及图像识别[J]. 计算机工程与科学, 2018, 40(10): 1858-1866.
[2]	王民，苏利博，王稚慧，要趁红. 采用STRAIGHT模型和深度信念网络的语音转换方法[J]. 计算机工程与科学, 2016, 38(09): 1950-1954.