基于Lighten CNN的自动人脸分类模型的优化方法  

Optimization method of automatic face classification model based on lighten CNN

作　　者：孙旭 胡伟[1] 李瑞瑞 SUN Xu;HU Wei;LI Ruirui(College of lnformation Sciences and Technology,Beijing University of Chemicial Technology,Beijing 100029,China)

机构地区：[1]北京化工大学信息科学与技术学院,北京100029

出　　处：《计算机应用》2018年第A01期32-35,82,共5页journal of Computer Applications

基　　金：国家自然科学基金资助项目(61501018;61571033)

摘　　要：随着多媒体、人工智能以及物联网等技术的飞速发展,对动态视频中的人脸的检测与分析是一个研究热点。针对目前对人脸特征分析的研究无法同时满足监控视频对准确性和实时性的需求,提出基于Lighten卷积神经网络(Lighten CNN)框架的单网络多目标的人脸特征分析方法。首先,提出模型训练样本拓充的半自动化数据标注策略;然后,结合多任务学习思想,基于Lighten CNN构建单网络多目标的网络结构;最后,在监控视频的人脸数据集上,进行人脸年龄、性别特征分析的实验,并基于人脸序列进行可信度分析。实验结果表明,所提出的基于Lighten CNN的人脸分类模型的优化方法能同时满足监控视频对特征分析的准确性和实时性的需求,具有良好的泛化能力和适用范围。With the rapid development of multimedia, artificial intelligence and Internet of things technology, the detection and analysis of human faces in dynamic video is a hot research topic. Concern the problem that the research on face feature analysis can＇ t meet the requirements of accuracy and real-time in surveillance video, a multi-objective face feature analysis method based on Lighten Convolutional Neural Network（ Lighten CNN） framework was proposed. Firstly, a semi- automatic data tagging strategy for model training and sample extension was proposed. Secondly, in order to combine the idea of multi-task learning, the single network and multi-target network structure was constructed based on Lighten CNN. Finally, the experiment of age and gender feature analysis was carried out on the dataset of surveillance video, and reliability analysis was done on the face sequence. The experimental results show that the proposed optimization method of face classification model based on Lighten CNN can simultaneously satisfy the needs of accuracy and real-time in surveillance video, and has good generalization ability and wide range of applications.

关 键 词：卷积神经网络 监控视频 人脸监测 特征分析 实时性 单网络多目标 人脸序列 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]