巩震 陈丹红



摘 要:目前,在普通手机解锁、面部锁定、面部扫描支付和安全防卫等领域,人脸识别信息技术都有着非常广泛的应用。传统的机器学习算法是基于人的脸部特征的,由于每类样本的不均衡性严重和场景效果的欠缺等因素,算法实现的效果常常不是很理想。本文将针对目前人脸识别技术领域的高语义特征及深度特征提取的缺陷,以及人工提取存在的误差,运用深度学习算法和类比中心等算法,提高人脸识别在特征提取领域的技术能力。利用卷积神经网络减少人工干预,提高特征提取的算法精度,从而提高损失函数的精确值。
关键词:人脸识别;算法优化;深度学习;卷积神经网络
中图分类号:TP312 文献标识码:A
文章编号:2096-1472(2021)-01-10-03
Abstract: Face recognition technology is widely used in ordinary mobile phone unlocking, face locking, face scanning payment, and security protection. Traditional machine learning algorithms are based on human facial features, and the effect is often not ideal due to serious imbalance of each type of sample and lack of scene effect. This paper aims to use deep learning algorithm and analog center algorithm to improve technical capability of face recognition with respect to feature extraction. Thus, defects of high semantic features and deep feature extraction, as well as errors in manual extraction will be greatly reduced. Convolutional neural network is used to reduce manual intervention, improve accuracy of feature extraction algorithm, so to improve the performance of loss function.
Keywords: face recognition; algorithm optimization; deep learning; convolutional neural network
1 引言(Introduction)
人脸识别始于20世纪60年代,随着计算机技术和光学成像技术在最近几年的兴起和神经网络技术的发展,特别是神经卷积网络的巨大成功,人脸识别系统在人的图像识别和检测脸型的效果方面有了很大的提高。但是,传统的人脸识别系统下算法分析技术的发展仍然存在巨大的不足,这也是本文想要探讨和改善的。对此,本文将利用人脸面部检测、局部二值模式、线性鉴别分析和深度学习网络来对传统的人脸识别流程做相关的优化和修改,促使计算机能够在最大程度上深入图像的高语义,进一步挖掘其深度特征。同时,针对对比损失算法、采样算法、类比中心损失算法、深度学习下Heatmap算法和轨迹追踪算法做进一步的修改,不断提高其精准度,减小其具体的误差。……