【作者】 徐熙君；

【导师】 许曰滨；

【作者基本信息】 青岛大学 ， 计算机科学与技术， 2005， 硕士

【摘要】 人脸识别技术以其潜在的应用广泛性,在过去的几十年中得到了迅速的发展。但是,由于人脸非刚性、表情多变等诸多因素,使得人脸识别在实际应用中面临巨大困难,人脸识别已经成为一项极具挑战性的研究课题。本文将遗传算法应用于人脸识别的图像分割、人脸定位和角度矫正等环节,建立了相应的数学模型。 具体做法是,在图像分割的遗传算法设计中,采用Otsu法中的最大类间方差和KSW法中的总熵作为适用度度量;在人脸的识别算法中,建立了人脸矩形区域模板中双眼、鼻子和嘴巴存在性度量检测数学模型,以它们存在性度量的加权和作为遗传算法的适用度度量函数。在遗传算法的算子和执行策略的设计中,采用“精英选择”策略、由父代和子代染色体参与的两点和多点杂交算子以及由两条染色体参与的变异算子,从而最大限度地避免了早熟收敛。另外,在遗传算法设计中还添加了一种“逆转算子”,即在一条染色体上随机选择两点构成一个子串,首尾倒置形成新子串替代原子串在个体中的位置。 本文对于提出的遗传算法,给出了遗传算子的数学描述,并应用遗传算法概率收敛性理论和马氏链理论对算法进行了收敛性分析,对提出的遗传算法的收敛性进行了证明。 仿真实验表明,本文提出的遗传算法与标准遗传算法相比,其算法的收敛速度、解的精度以及人脸识别率都有显著的提高。更多还原

【Abstract】 Owing to its potential and extensive applications, human face recognition technology has been developed rapidly in the passed decades. Because human face is nonrigid and its expression is changeful, human face recognition has been facing tremendous difficulties in actual applications, which makes human face recognition become a challenging issue. In this article, GAs are applied in the principal steps of face recognition, including image segmentation, location of face and curing the angle, and a mathematical model is built.The concrete means is to take maximum variance between two classes in Ostu method and total entropy in KSW method as fitness in image segmentation;in the detection and location of face, a mathematical model is established to measure the existence of two eyes, nose and mouse in a rectangle area, and weight-based sum of these measurements is adopted as GA’s fitness. In the design of GAs’ operators and performing strategies, selection operator is designed to be ’elite selection’;multi-point crossover can perform between father generation and son generation;mutation requires two chromosomes, thus premature is avoided by the greatest extent. In addition, an inversion operator is given to generate a new binary string by choosing randomly two gene positions and inversing the substring between the two positions in a chromosome.For the GAs given in this article, a mathematical description is presented and the GAs’ convergence properties are analyzed based on GA probability convergence theory and Markov chain theory, and the proposed algorithms are proved to be global convergent.The emulation illustrates that the convergence velocity, optimal solution of the presented genetic algorithms and the rate of face recognition, compared with the standard genetic algorithm, are enhanced significantly.更多还原