卷积神经网络用于人脸特征提取  

Facial feature extraction with CNN

作　　者：陈兵 蒋行国 CHEN Bing;JIANG Xingguo(Sichuan University of Science&Engineering,Yibin 644000,China;Sichuan Key Laboratory of Artificial Intelligence,Yibin 644000,China)

机构地区：[1]四川轻化工大学,四川宜宾644000 [2]人工智能四川省重点实验室,四川宜宾644000

出　　处：《现代电子技术》2022年第18期182-186,共5页Modern Electronics Technique

基　　金：人工智能四川省重点实验室基金(2020RZJ03);人工智能四川省重点实验室基金(2018RZJ01);四川轻化工大学人才引进项目(2019RC12);四川轻化工大学人才引进项目(2018RCL18)。

摘　　要：卷积神经网络作为深度学习的代表结构之一,在特征提取方面相较于传统方法具有更强的特征学习与特征表达能力。文中基于卷积神经网络在图像特征提取方面的优点,提出一种用于人脸特征提取的卷积神经网络模型。该模型以AR人脸数据集为实验数据,利用不同卷积层和LBP方法分别对人脸进行特征提取,并对提取特征进行分析;然后使用支持向量机SVM对提取特征进行预测分类,分析采用不同方法提取特征对分类结果的影响。实验结果表明,卷积神经网络提取特征分类准确率高于LBP方法,且第2卷积层准确率最高达到73.8%。在严格限定拍摄角度和光照的情况下,所提出的模型能够提取人脸深层特征,且人脸图像特征在表达性方面优于传统LBP方法,可为构建人脸数学模型提供参考价值。As one of the representative structures of deep learning,the CNN(convolutional neural networks)has stronger ability of feature learning and feature expression than traditional methods in feature extraction. A CNN model for face feature extraction is proposed based on its advantages in image feature extraction. In this model,AR facial database is taken as experimental data,different convolution layer and LBP(local binary pattern)methods are used to extract and analyze the facial features,the support vector machine(SVM) is used to predict and classify the extracted features,and then the influence of different extraction methods on classification results is analyzed. The experimental results show that the classification accuracy of feature extraction with CNN is higher than that of LBP method,and the accuracy of the second convolution layer is up to 73.8%.In the case of strictly limiting the shooting angle and illumination,the proposed model can extract deep facial features,and the facial image features are better than the traditional LBP method in expressivity,which provides a reference value for the construction of face mathematical model.

关 键 词：深度学习 卷积神经网络 人脸识别 特征选择 特征提取 图像分类 支持向量机 

分 类 号：TN926-34[电子电信—通信与信息系统]