【作者】 张超；

【导师】 赵健；

【作者基本信息】 西北大学 ， 信号与信息处理， 2018， 硕士

【摘要】 人脸识别作为计算机视觉以及人工智能领域中一个非常重要的研究方向,在基于人脸的疾病筛查以及结合大数据的交通智能监控等领域具有非常高的研究意义与应用价值。稀疏表示是近几年的热点之一,在信号处理领域具有鲁棒性强、精确等优点。将稀疏表示理论应用于人脸识别问题在理论上已经证明其可行性,并通过实验证实了其稳定性与高效性。本文针对人脸多姿态评估问题、姿态针对性数据降维问题和姿态鲁棒性稀疏字典构造方法等问题进行了深入的分析与研究,最终目的是提升算法在多姿态问题上的鲁棒性。其主要研究和创新内容如下所示:(1)对多姿态评估问题进行了详细的分析。根据面部主要器官的绝对位置与几何相对关系,提取并构建姿态特征向量,并结合随机森林算法进行人脸多姿态评估。最终通过实验仿真对所提方法进行验证。对人脸数据降维主要方法做了介绍,分析每一个像素点在姿态评估中所占比重,提出了一种具有姿态针对性的人脸数据降维方法。(2)针对多姿态人脸识别问题,在字典构造环节中,提出双字典算法(双字典分为正脸字典与侧脸字典)提升识别算法在应对多姿态问题时的鲁棒性。并与多姿态评估研究结果相结合,提出了基于随机森林与双字典稀疏表示的姿态鲁棒性人脸识别方法。在与同类单字典算法的对比中,所提方法在多姿态鲁棒性上得到了更好的效果,在FERET库中识别率平均提升9.90%,在Pointing Data库中识别率平均提升1.90%。更多还原

【Abstract】 Face recognition is a very important research field in the field of computer vision and artificial intelligence.Face recognition has very high research significance and application value in human face-based disease screening and traffic intelligence monitoring combined with big data.Sparse representation theory is one of the research hotspots in recent years.Sparse representation has the advantages of strong robustness and accuracy in signal processing.The theory of sparse representation applied to face recognition has proved its feasibility theoretically.Experiments show that sparse representation of face recognition stability and efficiency.This thesis analyzes and researches deeply on face multi-posture appraisal problems,dimension-specific data dimensionality reduction problems,and robust sparse dictionary construction methods.The ultimate goal is to improve the robustness of the algorithm on multi-pose problems.The main research and innovations of this thesis are as follows:(1)The issue of multi-pose assessment was analyzed in detail.According to the relative position of the main facial organs and the relative geometric relationship,the pose eigenvectors are extracted and constructed,and the random forest algorithm is used to evaluate the face multi-pose.Finally,the proposed method is verified through experimental simulation.The main method of dimension reduction of face data is introduced.By analyzing the proportion of each pixel in the facial pose evaluation,a face data dimension reduction method is proposed of posture targeted.(2)To solve the problem of multi-pose face recognition,this thesis proposes a dual-dictionary idea(dictionaries divided into full-face dictionary and profile dictionary)to improve the robustness of the recognition algorithm in solving multi-pose problems.Combined with the result of multi-pose assessment,a pose-robust face recognition method based on random forest and dual-dictionary sparse representation is proposed.Compared with similar algorithms,the proposed method achieves better results in multi-pose robustness.The recognition rate in FERET face database increases by 9.90% on average.The recognition rate in the Pointing Data face database increased by an average of 1.90%.更多还原