遮挡条件下的人脸识别算法研究

发布时间：2018-05-09 11:56

本文选题：人脸识别 + 二维Gabor　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：随着社会的发展,人脸识别在身份认证、人机交互、视频监控等方面得到了广泛应用。人脸识别技术在各行各业中发挥着巨大的作用,但是存在着很多难题有待攻克。本文主要研究遮挡条件下的人脸识别问题,从连续性遮挡的处理能力、人脸的特征提取、遮挡和非遮挡区域的图像分割三个方面入手,展开了一系列研究工作。本文主要贡献和创新点如下:(1)针对传统特征提取方法进行实验分析,本文通过特征向量选择实验分析了特征向量个数对人脸识别的影响。该实验反映了PCA算法忽略了不同样本之间的差异性,缺乏对局部特征信息的表示。另外,通过相似度匹配实验和人脸局部特征检测实验对二维Gabor函数进行分析,这两组实验说明了二维Gabor作为数学变换的核函数,可以很好的提取人脸的局部特征信息。(2)本文基于双属性模型改进PCA算法,然后融合局部的二维Gabor算法,提出了新的人脸识别算法(Double Attribute Model based Gabor,DAMG),从线性子空间角度研究解决遮挡条件下的人脸识别问题。该算法基于双属性模型将全局特征向量和误差特征向量融合生成双属性特征向量,根据二维Gabor对目标图像进行分块化特征提取,设计整体分类器对双属性特征向量和局部分块向量进行加权分类。本文通过DAMG算法的性能分析和识别错误的样本分析证明了该算法在遮挡人脸的识别过程中具有非常好的鲁棒性。(3)本文提出基于CV模型的双加权误差分布模型,从高维图像表示的角度研究解决遮挡条件下的人脸识别问题。首先,基于CV(Chan-Vese)模型对遮挡图像进行分割得到不同区域下的误差图像。其次,提出了基于梯度方向的条件概率误差模型,与低维度特征向量相比,条件概率误差模型具有更多的特征信息。最后,根据遮挡区域和非遮挡区域的图像误差分布,分别推算出两种误差的分布模型,形成双加权误差分布模型。本文通过人脸图像随机遮挡、人脸五官遮挡、真实人脸遮挡这三组实验验证了该算法和DAMG算法在遮挡条件下进行人脸识别的有效性。
[Abstract]:With the development of society, face recognition has been widely used in identity authentication, human-computer interaction, video surveillance and so on. Face recognition technology plays a great role in various industries, but there are many problems to be solved. In this paper, face recognition under occlusion conditions is mainly studied. A series of research work is carried out from three aspects: processing ability of continuous occlusion, feature extraction of face, image segmentation of occlusion and unoccluded region. The main contributions and innovations of this paper are as follows: (1) based on the experimental analysis of traditional feature extraction methods, this paper analyzes the effect of the number of feature vectors on face recognition through feature vector selection experiments. The experiment shows that the PCA algorithm ignores the differences between different samples and lacks the representation of local feature information. In addition, the two-dimensional Gabor function is analyzed by similarity matching experiment and face local feature detection experiment. The two groups of experiments show that two-dimensional Gabor is the kernel function of mathematical transformation. This paper improves the PCA algorithm based on the two-attribute model, and then fuses the local two-dimensional Gabor algorithm. A new face recognition algorithm, double Attribute Model based Gaboran, is proposed to solve the problem of face recognition under occlusion from the perspective of linear subspace. Based on the two-attribute model, the global eigenvector and the error eigenvector are fused to generate the two-attribute eigenvector, and the target image is extracted by block feature extraction based on two-dimensional Gabor. A global classifier is designed for weighted classification of dual attribute feature vectors and local block vectors. Through the performance analysis of DAMG algorithm and the sample analysis of error recognition, it is proved that the algorithm has very good robustness in the process of shading face recognition. In this paper, we propose a double-weighted error distribution model based on CV model. Face recognition under occlusion condition is studied from the point of view of high dimensional image representation. Firstly, the occlusion images are segmented based on CVV Chan-Vese model to obtain the error images in different regions. Secondly, the conditional probability error model based on gradient direction is proposed. Compared with the low-dimensional eigenvector, the conditional probability error model has more feature information. Finally, according to the image error distribution of occlusion region and unoccluded region, the distribution models of two kinds of errors are deduced, respectively, and the double-weighted error distribution model is formed. In this paper, the experiments of random occlusion, facial facial occlusion and real face occlusion in face images demonstrate the effectiveness of the proposed algorithm and DAMG algorithm for face recognition under occlusion conditions.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】