摘要
非限制场景下存在光照、遮挡和姿态变化等问题,这严重影响了人脸识别模型的性能和准确度。针对该问题,本文对facenet进行改进,提出了一种基于facenet皮尔森判别网络的人脸识别方法facenetPDN。首先,构建facenetPDN深度卷积神经网络,在facenet前端融合多任务级联卷积神经网络进行人脸检测提取目标人脸。然后,通过深度神经网络提取人脸深度特征信息,采用皮尔森相关系数判别模块替换facenet中的欧氏距离判别模块实现人脸深度特征判别。最后,使用CASIA-WebFace和CASIA-FaceV5人脸数据集训练网络。为了证明本文方法的有效性,训练后的模型在LFW和celeA人脸数据集进行测试和评估,并进行对比分析。实验结果表明,改进后的facenetPDN方法的准确度比原来整体提高了1.34%,在融合训练集下提高了0.78%,该算法鲁棒性和泛化能力优良,可实现多人种的人脸识别,对非限制场景下人脸目标具有良好的识别效果。
In unrestricted scenes,there are problems such as illumination,occlusion,and pose changes,which seriously affect the performance and accuracy of the face recognition model.This paper improves facenet to solve this problem,and proposes a facenetPDN method based on facenet Pearson discriminant network.Firstly,the deep convolutional neural network in facenetPDN is constructed,and the multi-task cascaded convolutional neural network is fused on the front-end of facenet to detect and extract the target face.Then,the facial depth feature information is extracted through the deep convolutional neural network,and the Pearson correlation coefficient discrimination module is used to replace the Euclidean distance discrimination module in the facenet algorithm to realize the facial depth feature discrimination.Finally,CASIA-WebFace and CASIA-FaceV5 face datasets are used to train the network.The trained model is tested and evaluated on the LFW and celeA face datasets to prove effectiveness of the method in this paper,and a comparative analysis is performed.The experimental results show that the accuracy of the improved facenetPDN method is 1.34%higher than that of the original method as a whole,and the accuracy of the model after training in the fusion training dataset is improved by 0.78%.The algorithm has excellent robustness and generalization ability,which can realize multi-ethnic face recognition,and has a good recognition effect on face targets in unrestricted scenes.
作者
谷凤伟
陆军
夏桂华
GU Fengwei;LU Jun;XIA Guihua(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China;Key Laboratory of Intelligent Technology and Application of Marine Equipment,Harbin Engineering University,Harbin,150001,China)
出处
《智能系统学报》
CSCD
北大核心
2022年第1期107-115,共9页
CAAI Transactions on Intelligent Systems
基金
黑龙江省自然科学基金项目(F201123).
作者简介
谷凤伟,博士研究生,主要研究方向为人脸识别、机器视觉、图像处理、目标跟踪;通信作者:陆军,教授,博士生导师,主要研究方向为计算机视觉、智能控制技术、图像处理、高性能船舶控制和机械臂控制。科技部科技型中小企业创新基金项目评审专家,国家自然科学基金同行评议专家。获省部级以上奖励多项。发表学术论文70余篇.E-mail:lujun0260@sina.com;夏桂华,教授、博士生导师,主要研究船舶运动建模与仿真、智能船舶技术、船舶驾控模拟系统、图像处理和智能机器人控制。享受国务院特殊津贴专家,获省部级以上奖励多项。发表学术论文70余篇。
