人脸识别中光照预处理算法研究
发布时间：2018-11-16 11:16
【摘要】：人脸识别,是生物特征识别领域中的热门研究话题,同时也是计算机视觉领域最成功的应用之一。它具有广泛的应用前景,在门禁系统,智能安防,智能监控以及国家军事和安全领域等表现出了无可替代的作用和潜力。通过将近五十年的研究,该技术目前已经取得了很大的进步,并且投入到商业使用。但是总体而言,人脸识别仍然存在一些难题,比如在光照条件差,用户配合度低的情况下,识别性能将会迅速降低。因此本文综合分析了其中一个因素即光照变化给人脸识别带来的影响,并对其深入研究,提出改进算法。本文详细分析了基于图像处理的基本算法与基于朗伯模型的光照不变量提取算法,并且根据理论分析结果提出了两种改进算法。本文提出的第一种算法是一种基于维纳滤波的自商图像算法。自商图像算法克服了传统商图像的局限性,通过自商模型可以提取与光照无关的内在特性,但是该算法对人脸表面光照的估计使用高斯滤波。这类算法没有考虑到该滤波器在平滑人脸表面的同时也将轮廓特征模糊化,而本文采用的自适应维纳滤波能够根据人脸表面局部方差值自动调整滤波强度,从而更好的获取人脸的本质特征表达。本文提出的第二种算法是基于多方向的相对梯度算法。文中通过理论验证得到图像的相对梯度特征也具有光照不变性,传统梯度图像的计算仅仅考虑X和Y两个方向的梯度分量,而本文综合考虑了四个方向,通过高斯函数一阶导数与图像作卷积计算出多个方向的梯度分量,然后对每个方向的相对梯度分量作加权融合,从而获得更好的人脸图像本质特征表达。本文提出的两种改进算法分别都在Extended Yale B和CMU PIE人脸库上做实验,同时本文的算法同多个光照处理算法作了对比,统一使用最近邻作为分类标准。大量的实验结果表明,本文提出的算法能够有效的获取光照不变分量,提高人脸识别准确率。
[Abstract]:Face recognition is a hot topic in biometric recognition field, and it is also one of the most successful applications in computer vision field. It has a wide application prospect, and has shown irreplaceable role and potential in access control system, intelligent security, intelligent monitoring and national military and security fields. Through nearly 50 years of research, the technology has made great progress and put into commercial use. But in general, there are still some problems in face recognition, such as poor lighting conditions and low user cooperation, the performance of face recognition will be reduced rapidly. Therefore, this paper comprehensively analyzes one of the factors, that is, the influence of illumination change on face recognition, and makes a thorough study of it, and proposes an improved algorithm. In this paper, the basic algorithm based on image processing and the illumination invariant extraction algorithm based on Lambert model are analyzed in detail, and two improved algorithms are proposed according to the results of theoretical analysis. The first algorithm proposed in this paper is a self-quotient image algorithm based on Wiener filter. The self-quotient image algorithm overcomes the limitation of the traditional quotient image and can extract the inherent characteristics independent of illumination through the self-quotient model. But Gao Si filter is used to estimate the illumination of the face surface. This algorithm does not take into account that the filter not only smoothes the face surface but also blurs the contour features, and the adaptive Wiener filter can automatically adjust the filtering intensity according to the local square difference of the face surface. In order to obtain a better expression of the essential features of the face. The second algorithm proposed in this paper is a multi-directional relative gradient algorithm. In this paper, the relative gradient features of the image are also shown to be illumination invariant by theoretical verification. The traditional gradient images only consider the gradient components in X and Y directions, and the four directions are considered in this paper. The gradient components of multiple directions are calculated by convolution of the first derivative of Gao Si function with the image, and then the relative gradient components of each direction are weighted and fused to obtain a better expression of the essential features of the face image. The two improved algorithms proposed in this paper are experimented on Extended Yale B and CMU PIE face database respectively. The proposed algorithm is compared with several illumination processing algorithms and the nearest neighbor is used as the classification standard. A large number of experimental results show that the proposed algorithm can effectively obtain illumination invariant components and improve the accuracy of face recognition.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
[Abstract]:Face recognition is a hot topic in biometric recognition field, and it is also one of the most successful applications in computer vision field. It has a wide application prospect, and has shown irreplaceable role and potential in access control system, intelligent security, intelligent monitoring and national military and security fields. Through nearly 50 years of research, the technology has made great progress and put into commercial use. But in general, there are still some problems in face recognition, such as poor lighting conditions and low user cooperation, the performance of face recognition will be reduced rapidly. Therefore, this paper comprehensively analyzes one of the factors, that is, the influence of illumination change on face recognition, and makes a thorough study of it, and proposes an improved algorithm. In this paper, the basic algorithm based on image processing and the illumination invariant extraction algorithm based on Lambert model are analyzed in detail, and two improved algorithms are proposed according to the results of theoretical analysis. The first algorithm proposed in this paper is a self-quotient image algorithm based on Wiener filter. The self-quotient image algorithm overcomes the limitation of the traditional quotient image and can extract the inherent characteristics independent of illumination through the self-quotient model. But Gao Si filter is used to estimate the illumination of the face surface. This algorithm does not take into account that the filter not only smoothes the face surface but also blurs the contour features, and the adaptive Wiener filter can automatically adjust the filtering intensity according to the local square difference of the face surface. In order to obtain a better expression of the essential features of the face. The second algorithm proposed in this paper is a multi-directional relative gradient algorithm. In this paper, the relative gradient features of the image are also shown to be illumination invariant by theoretical verification. The traditional gradient images only consider the gradient components in X and Y directions, and the four directions are considered in this paper. The gradient components of multiple directions are calculated by convolution of the first derivative of Gao Si function with the image, and then the relative gradient components of each direction are weighted and fused to obtain a better expression of the essential features of the face image. The two improved algorithms proposed in this paper are experimented on Extended Yale B and CMU PIE face database respectively. The proposed algorithm is compared with several illumination processing algorithms and the nearest neighbor is used as the classification standard. A large number of experimental results show that the proposed algorithm can effectively obtain illumination invariant components and improve the accuracy of face recognition.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前2条
1 蔺，
本文编号：2335366
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2335366.html
