面向人脸识别的特征提取技术应用研究

发布时间：2018-03-29 16:55

本文选题：特征生成　切入点：特征提取　出处：《东南大学》2016年博士论文

【摘要】：人脸识别技术是模式识别和计算机视觉领域的一个经久不衰的研究方向,作为一种重要的生物特征识别技术,在社会公共安全、监控、身份验证等日常生活中具有广泛的应用前景。人脸识别的直接目的在于利用人脸图像识别和验证个体身份,而真实场景下的人脸识别由于外部环境的光照变化,物体遮挡,人脸姿态变化和面部表情变化等因素,具有极大的挑战。典型人脸识别系统包括以下五个部分:人脸检测,人脸预处理,特征生成,特征提取/特征选择,匹配和识别。而人脸特征生成和特征提取对人脸识别的精度具有最直接的影响,因此主要关注人脸识别系统中的特征生成和特征提取环节,研究如何从原始图像中提取有效的面部特征及如何从高维面部特征中抽取具有鉴别能力的信息,从而提高人脸识别性能。本文分别研究了经典的局部特征描述子,稀疏低秩表示理论和协同表示方法,并在公开人脸数据集上进行了大量的实验进行验证方法的有效性。本文研究的具体内容包括:(1)针对局部三值模式这类特征描述子方法在特征生成中需要选取合适的阈值来克服不同噪声的问题,本文给出了一种阈值自适应的局部三值模式特征和中心对称自适应局部三值模式,该方法利用韦伯法则,自动根据像素的灰度值选择与之对应的阈值,从而解决固定阈值不能适应像素变化的缺陷,此外中心对称自适应局部三值模式比自适应局部三值模式具有更低的特征维数。在ORL和FERET人脸数据库实验表明,本文提出的两种方法的识别率均优于传统的局部特征描述子方法。(2)稀疏低秩表示要求字典是过完备的,故特征提取(维数约减)仍是重要的工作。本文首先利用低秩表示理论构建关联图,给出了一种基于低秩表示理论的特征提取方法,该方法利用低秩表示理论构建核范数图,并在此基础上刻画样本局部紧密度和总体离散度;其次研究利用降维后子空间内低秩表示关系设计原空间的关联关系,给出了一种两步迭代低秩表示投影的特征提取方法;最后,利用稀疏表示分类策略,给出一种低秩表示分析方法直接用于特征提取,避免构造图嵌入学习中的关联图。在FERET、AR、ORL人脸库和PolyU KFP指关节库上的实验表明了上述方法的有效性。(3)利用协同表示分类识别效果好,运算速度快的优点,本文给出了一种协同表示投影分析特征提取方法,有效丰富了图嵌入学习框架。该方法基于L2范数图构建刻画样本局部精密度和总体离散度,根据Fisher鉴别分析思想建立目标函数,利用广义特征值分解计算投影矩阵。进一步,本文给出一种非线性核协同表示分类方法,有效增强了协同表示分类方法的性能。该方法利用核技巧,将原始不可分的特征空间映射到高维可分的特征空间,进行优化求解。在多个公开数据集上的实验结果表明,本文提出的方法明显优于经典的方法。(4)为了识别远距离监控系统低分辨率人脸图像,将高分辨率图片和低分辨率图片分别看作两组不同的变量,利用典型相关分析理论,将他们投影到同一个线性空间中,实现不同分辨率图片的匹配。利用此方法,本文提出了一种基于典型相关分析的远距离低分辨率退化人脸识别方法,能有效克服分辨率不一致和分辨率低的问题。在Extended Yale B、ORL和AR人脸库上的实验结果表明本文方法对低分辨率图像具有较好的鲁棒性。
[Abstract]:The technology of face recognition is an important research direction and field of pattern recognition and computer vision is not bad, as a kind of important biometric technology, monitoring in public security, and has broad application prospects such as identity authentication in daily life. The direct purpose of face recognition using face recognition and verification of identity however, face recognition in real scene due to external environment changes in illumination, object occlusion, face posture and facial expression changes and other factors, is a great challenge. The typical face recognition system includes five parts: face detection, image preprocessing, feature generation, feature extraction and feature selection, matching and recognition while facial feature generation and feature extraction for face recognition accuracy is the most direct impact, so the feature generation and feature extraction ring focus in face recognition system Day, study how to effectively extract facial features from the original image and how to extract facial features from high dimension has ability to identify information, so as to improve the performance of face recognition. This paper studies the local features of classical, low rank sparse representation theory and collaborative representation method, and the validation method of the experiments were performed in the open face data sets. The specific contents of this paper include: (1) according to the local three value pattern of this feature descriptor in the feature generation need to select the appropriate threshold to overcome different noise problems, this paper presents an adaptive threshold value of three local and central symmetry adaptive pattern local three value model, the method of using Weber's law, automatically select the corresponding threshold according to the gray value of pixels, so as to solve the fixed threshold can not adapt to the change of pixel In addition, defects, central symmetry adaptive local three value model than the adaptive local three value model has a lower dimension. In the ORL and FERET face database show that the recognition method of local feature descriptor, this paper proposes two methods were better than traditional. (2) low rank sparse overcomplete dictionary is required, so feature extraction (dimensionality reduction) is an important work. This paper use low rank representation theory to construct the graph, is presented based on low rank representation theory of feature extraction method, the method of using the low rank representation of nuclear norm graph theory construction, and based on the characterization of local sample compactness and overall dispersion; secondly Study on using low dimensional subspace in low rank representation of the original space relationship between design, gives a two step iterative low rank representation extraction method of projection features; finally, using sparse representation points Such strategies, given a low rank representation of the direct analysis method for feature extraction, avoid association graph structure graph embedding learning. In FERET, AR, ORL face database and PolyU KFP refers to the joint library. Experimental results show that this method is effective. (3) the use of collaborative representation classification and recognition effect is good, the advantages of operation fast speed, this paper presents a collaborative representation projection analysis method of feature extraction, effectively enrich the graph embedding framework. The learning method based on L2 norm maps depict local sample precision and overall dispersion, according to the Fisher identification analysis thought establish objective function, using the generalized eigenvalue decomposition of the projection matrix is computed. Further, this paper given a nonlinear kernel collaborative representation classification method, effectively enhance the performance of collaborative representation classification method. This method uses the kernel feature space mapping technique, the original can not be separated into high dimensional points. Eigen space optimization algorithm. In a number of public data sets. The experimental results show that this new method is superior to the classic. (4) in order to identify the remote monitoring system of low resolution face image, the high resolution images and low resolution images are regarded as two different groups of variables, using the theory of canonical correlation analysis. They will be projected to the same linear space, realize the matching of different resolution images. By using this method, this paper proposes a face recognition method of remote degraded low resolution based on canonical correlation analysis, can effectively overcome the inconsistent resolution and low resolution of the problem. In the Extended Yale B, ORL and AR face image databases. The experimental results show that this method is robust to the low resolution image.

【学位授予单位】：东南大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】