人脸识别背后的数据清理问题研究  

Research on the data cleansing problem for face recognition technology

作　　者：夏洋洋 龚勋[1] 洪西进[1,2] 

机构地区：[1]西南交通大学信息科学与技术学院,四川成都611756 [2]台湾科技大学资讯工程系,台湾台北10607

出　　处：《智能系统学报》2017年第5期616-623,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61202191);计算智能重庆市重点实验室开放基金项目(CQ-LCI-2013-06);国家重点研发计划项目(2016YFC0802209)

摘　　要：人脸识别技术在深度卷积神经网络(deep convolution neural networks,DCNN)的快速发展下取得了显著的成就。这些成果主要体现在更深层次的DCNN架构和更大的训练数据库。然而,由大多数私人公司持有的大型数据库(百万级)并不对外公开,即使当前部分开放的大型数据库,因为标注信息过少,无法保证精度,会影响DCNN的训练。本文提出了一种易于使用的多角度清理图像方法来提高数据的准确性:通过人脸检测算法清除掉无法检测到人脸的图像;在清理后的数据集上利用已有模型提取图像特征,并计算相似度,进而统计出一类人脸图像中每一张图像与其他图像不相似的数目,根据改进参数清理数据。实验表明,清理后的数据库训练模型在LFW和Youtube Face数据集上测试的精度得到了提升,使用较小规模数据集情况下,在LFW数据集上取得了99.17%的准确率,在Youtube Face数据集也达到了93.53%的准确率。Face recognition technology has made a significant progress in the rapid development of deep convolution neural networks（ DCNN）. These developments are mainly focused toward a denser DCNN architecture and larger training database. However,DCNN training is affected because the large-scale database held by most private companies are not publically accessible. Moreover,current large-scale open databases are not accessible because of the slight availability of the labeled information and hard-to-guarantee accuracy. This study presents an easy-to-use image cleansing method to improve the accuracy of data from the following perspectives： First,deleting the face image that cannot be detected by face detection; second,using the existing model to extract the features of an image on the cleaned dataset and calculate the similarity; and finally,counting the number of images that are unlike the other images. The data were cleansed according to the improved parameters extracted from the abovementioned perspectives. The experimental results reveal that the cleansed database training model has improved the accuracy of face recognition in LFW（ labeled faces in the wild） and You Tube face database. In the case of using a small-scale dataset,an accuracy of 99.17% and 93.53% was achieved on the LFW and You Tube face datasets,respectively.

关 键 词：深度卷积神经网络 DCNN 清理图像 人脸识别 大型数据库 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]