基于深度学习的非限定条件下人脸识别研究

发布时间：2018-03-03 13:30

本文选题：人脸识别　切入点：数据清理　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来随着深度卷积神经网络(deep convolution neural networks,DCNN)的引入,人脸识别的准确率得以跨越式提升,各类相关应用如人脸识别考勤,考生身份验证,刷脸支付,人脸归类查询等已开始逐步投入使用,效果显著。然而现实场景中非限定条件的人脸识别技术却存在较多难题,例如姿态、光照、遮挡等困难,识别精度也随着数据规模的增加和识别难度的增大而快速下降。目前的解决方案是通过增大复杂场景下的训练数据库规模来学习到尽可能多的场景。然而大多数大型的数据库(百万级)由私人公司持有且不公开,即使目前公开的大型数据库,也由于标注信息过少、准确性得不到较好保证,会影响DCNN的训练。本文从确保训练数据库的准确性和解决部分姿态问题出发,主要做出了以下工作:本文首先深入了解了深度学习的部分理论知识,总结了当前主流的深度学习开源框架和人脸识别开源项目以及国内当前主流人脸识别相关公司和解决方案。通过检测误差、检测效果和效率以及在大规模数据上的稳定性的实验对比分析了当前主流的人脸检测算法的性能,并根据实际情况提出了检测评价机制减少误差,最终选择针对大规模数据进行处理综合效果好的算法,提出了基于关键点映射的人脸图像归一化算法。针对大规模数据库准确性无法保证、存在噪声等问题,提出了基于多角度评价的数据清理方法,在同一个类别中每张图像与其他图像进行相似度计算,并统计与该图像不相似的图像数量,超过一定的数量就对该图像进行清理。通过多方面的实验验证了清理数据方法的有效性。实验证明,清理后的数据库训练模型在LFW数据集上的准确率得到了提升,以较小规模的训练集取得了 99.17%的准确率,在Youtube数据集取得了 93.54%的准确率。为了解决非限定人脸识别中的多姿态问题,使用基于三维人脸模型的图像校正生成正面图像,并提出了将正面合成图像提取的特征与原始图像进行特征线性融合的方法来生成新的特征向量。合成正面图像能提供原始图像不具备的特征,也存在信息丢失,而原始图像的特征具有较高的参考价值,因此融合二者特征的新特征具有更全面的特征。实验证明,新的特征向量能够有效提高人脸识别率,将原有的LFW上50对错误匹配对矫正了 15对,在SWJTU-MF DB上也取得了显著的效果。
[Abstract]:In recent years, with the introduction of deep convolution neural networks (DCNN), the accuracy of face recognition has been improved by leaps and bounds. Face classification and query have been gradually put into use, and the effect is remarkable. However, there are many difficult problems in the face recognition technology, such as pose, illumination, occlusion and so on. Recognition accuracy also decreases rapidly with the increase of data size and difficulty. The current solution is to learn as many scenarios as possible by increasing the size of the training database in complex scenarios. Large databases (million levels) are held by private companies and are not publicly available, Even if the large database is open at present, the accuracy can not be guaranteed well because of too little tagged information, which will affect the training of DCNN. This paper starts from ensuring the accuracy of the training database and solving some posture problems. The main work is as follows: first of all, this paper has a deep understanding of some of the theoretical knowledge of in-depth learning, Summarizes the current mainstream deep learning open source framework and face recognition open source projects, as well as domestic mainstream face recognition related companies and solutions. The performance of the current mainstream face detection algorithms is compared and analyzed in the experiments of detection effect, efficiency and stability on large scale data. According to the actual situation, the detection evaluation mechanism is proposed to reduce the error. Finally, an algorithm for processing large scale data is selected, and a face image normalization algorithm based on key point mapping is proposed. The accuracy of large-scale database can not be guaranteed, and there are some problems, such as noise, etc. A data cleaning method based on multi-angle evaluation is proposed, in which the similarity between each image and other images in the same category is calculated, and the number of images that are not similar to the image is counted. More than a certain number of images are cleaned. The validity of the data cleaning method is verified by experiments in many aspects. The experimental results show that the accuracy of the database training model after cleaning is improved on the LFW dataset. In order to solve the problem of multi-pose in unqualified face recognition, the accuracy rate of 99.17% is obtained with the smaller training set and 93.54% with the Youtube dataset. In order to solve the problem of multi-pose in unqualified face recognition, the image correction based on 3D face model is used to generate the frontal image. A new feature vector is generated by linear fusion of features extracted from frontal composite image and original image. The synthesized frontal image can provide features that the original image does not possess, and there is also information loss. However, the features of the original image are of high reference value, so the new features with the fusion of the two features have more comprehensive features. Experiments show that the new feature vector can effectively improve the face recognition rate. The 50 pairs of error matching pairs on the original LFW were corrected by 15 pairs, and a remarkable effect was obtained on SWJTU-MF DB.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】