人脸识别背后的数据清理问题研究

发布时间：2018-10-19 13:55

【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。
[Abstract]:Face recognition technology has made remarkable achievements with the rapid development of deep convolution neural network (deep convolution neural networks,DCNN). These results are mainly reflected in a deeper DCNN architecture and a larger training database. However, large databases (millions of levels) held by most private companies are not open to the public, even if the currently partially open large databases have too little information to guarantee accuracy and affect DCNN training. In this paper, an easy-to-use multi-angle image cleaning method is proposed to improve the accuracy of the data. The face detection algorithm is used to remove the image that can not detect the face, and the existing models are used to extract the image features from the cleaned data set. The similarity degree is calculated, and the number of dissimilar images between each image and other images in a human face image is calculated, and the data is cleaned according to the improved parameters. Experimental results show that the accuracy of the training model is improved on LFW and Youtube Face datasets, and 99.17% accuracy is achieved on LFW data sets using smaller data sets. The accuracy of Youtube Face data set is 93.53%.
【作者单位】：西南交通大学信息科学与技术学院;台湾科技大学资讯工程系;
【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)
【分类号】：TP183;TP391.41

【相似文献】