一种仿生的人脸不变特征提取方法  

Invariant facial feature extraction method with biologically-like mechanism

作　　者：杜兴[1] 龚卫国[1] 张睿[1] 

机构地区：[1]重庆大学光电技术及系统教育部重点实验室,重庆400044

出　　处：《仪器仪表学报》2011年第4期813-818,共6页Chinese Journal of Scientific Instrument

基　　金：国家863计划(2007AA01Z423);重庆市科技攻关计划(CSTC2009AB0175);公安部应用创新项目(2010YYCXC.QSJ074);中央高校基本科研业务费(CDJXS10122218);重庆市自然科学基金(CSTC2010BB2230);高等学校博士学科点专项科研基金(20100191120012)资助项目

摘　　要：为了提高基于子空间算法的人脸识别的识别率,提出一种仿生的人脸不变特征提取方法。通过模拟初级视皮层（V1）的信息处理机制,构建一个二层结构的分层网络提取人脸图像的不变特征。网络的第1层模拟V1简单细胞的功能,通过稀疏编码方法学习获得一组类似V1简单细胞的滤波器,利用该组滤波器提取图像的光照不变特征;第2层模拟V1复杂细胞的功能,通过局部极大值运算对第1层的输出在空间和尺度邻域内进行合并,得到对光照、表情、轻微姿态变化和面部局部细节变化具有鲁棒性的人脸不变特征。以此不变特征代替原始人脸图像作为子空间算法的输入,从而提高识别率。在FERET和ORL人脸库上的实验表明,相对于直接使用子空间算法,方法将识别率提高了4.95%~20.35%。A biologically-like invariant facial feature extraction method is proposed to improve the face recognition rate obtained using the methods based on subspace algorithms.A hierarchical network with two layers,which is constructed according to the information processing procedure in the primary visual cortex（V1）,is put forward to extract invariant feature from the face image.The first layer of the network,which simulates the function of the V1 simple cells,learns and obtains a group of V1-simple-cell-like filters using sparse coding method,and employs these filters to extract a set of illumination insensitive features from the face image.The second layer,which simulates the function of the V1 complex cells,merges the output of the first layer in neighborhood of positions and scales using local maximum operation,so as to obtain the facial feature robust to illumination,expression,slight pose change and local facial detail variations.The obtained invariant features are used to replace original face image as the input of a subspace algorithm,and the performance of face recognition is improved.Experiments on the FERET and ORL face databases show that compared with directly applying subspace algorithms to the image,using the proposed method can increase the recognition rate by 4.95% ~ 20.35%.

关 键 词：人脸识别 人脸不变特征 分层网络 稀疏编码 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]