摘要
本文针对传统的基于稀疏表示的人脸识别方法在小样本情况下对类内变化鲁棒性不强的问题,从特征的层面入手,提出了基于深度学习特征的稀疏表示的人脸识别方法。本方法首先利用深度卷积神经网络提取对类内变化不敏感的人脸特征,然后通过稀疏表示对所得人脸特征进行表达分类。本文通过实验,说明了深度学习得到的特征也具有一定的子空间特性,符合基于稀疏表示的人脸识别方法对于子空间的假设条件。实验证明,基于深度学习特征的稀疏表示的人脸识别方法具有较好的识别准确度,对类内变化具有很好的鲁棒性,特别在小样本问题中具有尤为突出的优势。
Focusing on the problems that the traditional sparse representation based face recognition methods are not quite robust to intra-class variations, a novel Sparse Representation via Deep Learning Features based Classification ( SRDLFC) method is proposed in this paper, employing a deep convolutional neural network to extract facial features and a sparse representation based framework to make classification. Experimental results in this paper also verifies the features extracted from deep convolutional network do satisfy the linear subspace assumption. The proposed SRDLFC proves to be quite effective and be robust to intra-class variations especially for under-sampled face recog-nition problems.
出处
《智能系统学报》
CSCD
北大核心
2016年第3期279-286,共8页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(61333015)
国家重点基础研究发展计划(2011CB302400)
关键词
机器学习
生物特征识别
深度学习
特征学习
子空间
小样本
稀疏表示
人脸识别
machine learning
biometric recognition
deep learning
feature learning
subspace
under-sampled recognition
sparse representation
face recognition