局部特征映射与融合网络的人脸识别优化算法  

Face Recognition Optimization Algorithm Based on Local Feature Mapping and Fusion Network

作　　者：徐武[1] 陈盈君 汤弘毅 杨昊东 秦浩然 XU Wu;CHEN Yingjun;TANG Hongyi;YANG Haodong;QIN Haoran(School of Electrical Information Engineering,Yunnan Minzu University,Kunming 650000,China;China Petroleum Transportation Company,Urumqi 830014,China)

机构地区：[1]云南民族大学电气信息工程学院,云南昆明650000 [2]中国石油运输公司,新疆乌鲁木齐830014

出　　处：《河南科技大学学报（自然科学版）》2023年第2期59-64,72,共7页Journal of Henan University of Science And Technology:Natural Science

基　　金：国家自然科学基金项目(U1802271);云南省民委项目(云南民族传统手工艺数字保护-以傈僳族制弩为例)。

摘　　要：针对传统特征提取算法的局限性,提出基于深度神经网络DeepLab v2的人脸识别改进算法。首先,对图像中人脸进行定位,采用DeepLab v2改进网络提取人脸的面部特征,通过加入压缩激励(SE)模块细化多角度纹理特征。其次,采用局部二值模式(LBP)特征映射对目标图像进行补充特征提取,细化纹理结构并减少光照噪声的干扰,提升识别的鲁棒性。最后,进行特征信息融合,采用分类模块对融合特征识别并分类处理。结果表明:对比经典目标检测算法YOLOv1和传统DeepLab算法,改进算法识别出多角度的人脸局部特征,且在正常光照下改进算法的识别精确度分别提高了3.1%和5.9%,在强光照下改进算法的识别精确度分别提高了9.5%和13.6%。In view of the limitations of traditional feature extraction algorithms,an improved face recognition algorithm based on deep neural network(DeepLab v2)was proposed.Firstly,the face in the image was locate,the DeepLab v2 improved network was used to extract the facial features of the face,and the multiangle texture features was refined by adding the squeeze excifation(SE)module.Secondly,the local binary patterns(LBP)feature map was used to supplement the feature extraction of the target image,to refine the texture structure and reduce the interference of illumination noise.The robustness of recognition was improved.Finally,feature information fusion was performed,and the classification module was used to identify and classify the fused features.The results show that by comparing the classical object detection algorithm(YOLOv1)and the traditional DeepLab algorithm,the improved algorithm can identify local features of faces from multiple angles.The recognition accuracy of the improved algorithm under normal lighting is improved by 3.1%and 5.9%respectively,and the recognition accuracy of the improved algorithm is improved by 9.5%and 13.6%respectively under strong light.

关 键 词：人脸识别 神经网络 SE模块 局部二值模式 softmax分类 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]