基于多特征融合CNN的人脸识别算法研究
|
摘要
针对传统卷积网络结构或使用单一特征融合方法进行人脸识别存在特征提取不全、训练准确率低的问题,提出一种基于多特征融合卷积神经网络(CNN)的人脸识别算法。算法在LeNet5结构基础上,通过不同卷积层映射图融合、多个特征核映射图融合和浅层纹理的轮廓特征和深层高级特征相融合,增加输出通道数,进而提高图像整体的语义信息。针对relu激活函数在神经网络训练过程中丢失负轴信息的缺陷,算法引入leaky relu激活函数。通过在ORL、AR、FERET三个人脸库上分别对LeNet5卷积神经网络结构、单一特征融合与算法进行对比实验。实验结果表明:算法通过多特征融合提取的特征信息更广,识别准确率高于单一特征融合,引入leaky relu激活函数后,网络收敛效果更好,同时,对遮挡、光照等干扰具有鲁棒性。
This paper proposed a convolutional neural network face recognition algorithm based on multi-feature fusion to overcome the problems of incomplete feature extraction and low training accuracy in face recognition by traditional convolutional network structure or single feature fusion method.On the basis of LeNet5 structure,this algorithm increases the number of output channels and improves the semantic information of the image as a whole through the fusion of different convolutional layer map,multiple feature core map,contour feature of shallow texture and advanced feature of deep texture.Aiming at the defect that relu activation function loses negative axis information in neural network training,this algorithm introduces leaky relu activation function.LeNet5 convolutional neural network structure,single feature fusion and this algorithm were compared and tested on ORL,AR and FERET face database respectively.The experimental results show that this algorithm extracts feature information more broadly by multi-feature fusion,and its recognition accuracy is higher than that of single feature fusion.When leaky relu activation function is introduced,the network convergence effect is better,and at the same time,it is robust to interference such as occlusion and illumination.
This paper proposed a convolutional neural network face recognition algorithm based on multi-feature fusion to overcome the problems of incomplete feature extraction and low training accuracy in face recognition by traditional convolutional network structure or single feature fusion method.On the basis of LeNet5 structure,this algorithm increases the number of output channels and improves the semantic information of the image as a whole through the fusion of different convolutional layer map,multiple feature core map,contour feature of shallow texture and advanced feature of deep texture.Aiming at the defect that relu activation function loses negative axis information in neural network training,this algorithm introduces leaky relu activation function.LeNet5 convolutional neural network structure,single feature fusion and this algorithm were compared and tested on ORL,AR and FERET face database respectively.The experimental results show that this algorithm extracts feature information more broadly by multi-feature fusion,and its recognition accuracy is higher than that of single feature fusion.When leaky relu activation function is introduced,the network convergence effect is better,and at the same time,it is robust to interference such as occlusion and illumination.
引文
[1] 胡敏,程天梅,王晓华.融合全局和局部特征的人脸识别[J].电子测量与仪器学报,2013,27(9):817-822.
[2] Turk M A,Pentland A P.Face Recognition Using Eigenface[J].International Journal of Computer Applications,2015,118(5):12-16.
[3] Belhumeur P N,Hespanha J P,Kriegman D J.Eigenfaces vs.Fisherfaces:Recognition using Class Specific Linear Projection[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,1997,19(7):711-720.
[4] Pervaiz A Z.Real Time Face Recognition System based on EBGM Framework[C].Cambridge,UK:12th International Conference on Computer Modelling and Simulation IEEE,2010:262-266.
[5] Sun Y,Wang X,Tang X.Deep Learning Face Representation from Predicting 10 000 Classes[C].Columbus,OH,USA:Proceedings of IEEE Conference on Computer Vision and Pattern Recognition(CVPR),2014:1891-1898.
[6] Taigman Y,Yang M,Ranzato M A,et al.DeepFace:Closing the Gap to Human- level Performance in Face Verification[C].Columbus,OH,USA:Proceedings of IEEE Conference on Computer Vision and Pattern Recognition(CVPR),2014:1701-1708.
[7] 石学超,周亚同,池越.基于多特征融合可调监督函数卷积神经网络的人脸性别识别[J/OL].[2018-02-07].http://www.arocmag.com/article/02-2019-04-012.html.
[8] 吴尧,邱卫根.基于改进的深度卷积神经网络的人脸识别[J].计算机工程与设计,2017,38(8):2246-2250.
[9] 王功鹏,段萌,牛常勇.基于卷积神经网络的随机梯度下降算法[J].计算机工程与设计,2018,39(2):441-445.
[10] Mass A L,Hannun A Y,Ng A Y.Rectifier Nonlinearities Improve Neural Network Acoustic Models[C].Atlanta,Georgia,USA:Proceedings of the 30th International Conference on Computer Vision.IEEE,2015:1026-1034.
[11] 周飞燕,金林,董军.卷积神经网络研究综述[J].计算机学报,2017,40(6):1229-1251.
[12] 常亮,邓小明,周明全,等.图像理解中的卷积神经网络[J].自动化学报,2016,42(9):1300-1312.
[13] 陈耀丹,王连明.基于卷积神经网络的人脸识别方法[J].东北师大学报(自然学科版),2016,48(2):70-76.
[14] Christian Szegedy,Vincent Vanhoucke,Sergey Ioffe,et al.Rethinking the Inception Architecture for Computer Vision[C].Las Vegas,NV,USA:IEEE Conference on Computer Vision and Pattern Recognition (CVPR),2016:2818-2826.
[2] Turk M A,Pentland A P.Face Recognition Using Eigenface[J].International Journal of Computer Applications,2015,118(5):12-16.
[3] Belhumeur P N,Hespanha J P,Kriegman D J.Eigenfaces vs.Fisherfaces:Recognition using Class Specific Linear Projection[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,1997,19(7):711-720.
[4] Pervaiz A Z.Real Time Face Recognition System based on EBGM Framework[C].Cambridge,UK:12th International Conference on Computer Modelling and Simulation IEEE,2010:262-266.
[5] Sun Y,Wang X,Tang X.Deep Learning Face Representation from Predicting 10 000 Classes[C].Columbus,OH,USA:Proceedings of IEEE Conference on Computer Vision and Pattern Recognition(CVPR),2014:1891-1898.
[6] Taigman Y,Yang M,Ranzato M A,et al.DeepFace:Closing the Gap to Human- level Performance in Face Verification[C].Columbus,OH,USA:Proceedings of IEEE Conference on Computer Vision and Pattern Recognition(CVPR),2014:1701-1708.
[7] 石学超,周亚同,池越.基于多特征融合可调监督函数卷积神经网络的人脸性别识别[J/OL].[2018-02-07].http://www.arocmag.com/article/02-2019-04-012.html.
[8] 吴尧,邱卫根.基于改进的深度卷积神经网络的人脸识别[J].计算机工程与设计,2017,38(8):2246-2250.
[9] 王功鹏,段萌,牛常勇.基于卷积神经网络的随机梯度下降算法[J].计算机工程与设计,2018,39(2):441-445.
[10] Mass A L,Hannun A Y,Ng A Y.Rectifier Nonlinearities Improve Neural Network Acoustic Models[C].Atlanta,Georgia,USA:Proceedings of the 30th International Conference on Computer Vision.IEEE,2015:1026-1034.
[11] 周飞燕,金林,董军.卷积神经网络研究综述[J].计算机学报,2017,40(6):1229-1251.
[12] 常亮,邓小明,周明全,等.图像理解中的卷积神经网络[J].自动化学报,2016,42(9):1300-1312.
[13] 陈耀丹,王连明.基于卷积神经网络的人脸识别方法[J].东北师大学报(自然学科版),2016,48(2):70-76.
[14] Christian Szegedy,Vincent Vanhoucke,Sergey Ioffe,et al.Rethinking the Inception Architecture for Computer Vision[C].Las Vegas,NV,USA:IEEE Conference on Computer Vision and Pattern Recognition (CVPR),2016:2818-2826.