基于孪生神经网络的小样本人脸识别  

Small sample face recognition based on siamese network

作　　者：万立志 张运楚[1,2] 葛浙东 王超[1] WAN Lizhi;ZHANG Yunchu;GE Zhedong;WANG Chao(School of Information and Electrical Engineering,Shandong Jianzhu University,Jinan 250101,China;Shandong Provincial Key Laboratory of Intelligent Building Technology,Jinan 250101,China)

机构地区：[1]山东建筑大学信息与电气工程学院,山东济南250101 [2]山东省智能建筑技术重点实验室,山东济南250101

出　　处：《山东建筑大学学报》2022年第1期79-85,99,共8页Journal of Shandong Jianzhu University

基　　金：国家自然科学基金青年科学基金项目(61503219)。

摘　　要：在基于深度学习的人脸识别领域中,孪生神经网络是一种解决过少样本数据降低模型性能并导致过拟合现象的有效方法。文章基于孪生神经网络提出一种引入CBAM混合域注意力机制的小样本人脸识别方法,对比实验将输出值映射为128维特征向量;采用对比损失函数,通过比较样本网络输出特征向量之间的欧氏距离来判断人脸的相似度。结果表明:在使用更少的训练数据情况下,孪生神经网络引入CBAM能够为不同通道的特征图设置不同的权重,可以显著提高模型的准确率;以实验准确率为性能指标,与其他4种人脸识别模型对比,文章所提出的算法对于小样本集的人脸识别准确率达到了98.12%。In the field of face recognition based on deep learning,siamese neural network is an effective method to solve the problem of over-fitting,which reduces the performance of the model with too few sample data.Based on the siamese neural network,this paper proposes a small sample face recognition method that introduces the CBAM hybrid domain attention mechanism.In comparative experiments,the output value is mapped to 128-dimensional feature vectors.The contrast loss function is used to compare the euclidean distance between the output feature vectors of the sample network in order to judge the similarity of human faces.The results show that when using less training data,the introduction of CBAM can set different weights for the feature maps of different channels,which can significantly improve the accuracy of the model.Taking the experimental accuracy as the performence indicator,compared with the other 4 types of face recognition,the algorithm proposed in the article achieves an accuracy of 98.12%for face recognition in a small sample set.

关 键 词：孪生神经网络 注意力机制 人脸识别 小样本 深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]