【作者】 宋宇；

【导师】 周激流；

【作者基本信息】 四川大学 ， 通信与信息系统， 2005， 硕士

【摘要】 统计学习理论是机器学习领域的最新进展。该理论的核心思想是通过控制学习机器的复杂度,实现对推广能力的控制。支持向量机(Support Vector Machines,简称SVM)是在统计学习理论的基础上发展起来的一种新的学习机器,其优点在于:由于采用了结构风险最小化原则代替经验风险最小化原则,较好的解决了小样本学习的问题;将低维的原始空间映射到高维的特征空间,把非线性问题转化为了线性问题,同时又采用了核函数的方法,巧妙的避开了高维空间的复杂运算,使算法的实现成为可能。正是因为这些优越性能,该技术己成为机器学习理论的研究热点,并在很多领域都得到了成功的应用,但是SVM也存在着一些比较明显的缺陷。例如:核函数的选择和参数的优化缺乏理论指导、训练算法的完善、缺乏和先验知识的综合能力等等。这些问题的存在,使得SVM的研究还需进一步的完善。 本文将支持向量机与人脸检测技术相结合,以核函数的选择和参数的优化为核心,对彩色图像的正面人脸检测做了一些有益的研究和探索。 本文的主要工作包括: (1) 提出基于遗传算法的核函数参数优化的方法,并用实验进行验证。 本文以Gauss核函数为例分析了核函数的参数选择对核函数性能的影响,并在此基础上提出了利用遗传算法的全局优化能力来解决核函数的参数选择问题,实验结果证明经该方法得出的最优参数取得了较好的效果,并且其搜索最优参数的速度较快。在此基础上本文利用遗传算法,以检测错误率为适应度,对Gauss核函数、3次多项式核函数、4次多项式核函数三种核函数进行了参数优化的对比实验,并仔细分析结果,我们发现:Gauss核函数具有较高的分类更多还原

【Abstract】 Statistical learning theory is a kind of new technology in machine learning field. Its main idea is to control learning machine’s generalization ability by controlling its model complexity. Support vector machine(SVM) is a new kind of learning machine based on statistical learning theory, which has many advantages. It solves small-sample problems by using structural risk minimization(SRM) to take the place of empirical risk minimization(ERM). Moreover, nonlinear problems are changed into linear ones by using mapping the low dimension original space to high dimension feature space, and employing kernel function, which make the algorithm be realized easily. Because of such advantages, SVM becomes a hot spot of machine learning theory, and is applied successfully. As a new technology, there are also many shortcomings that need to be researched, such as the adaptive kernel function and parameter selection, the limitation in the scale of training set, the shortcomings of training methods, and the combination with the prior knowledge, etc. Therefore, there is a need to research further.Based on the combination of SVM and face detection, research about frontal face detection in color images is discussed in this paper. Furthermore, the selection of kernel function and kernel parameters optimization are emphases in this paper. The main research work is as follows: (1) The method to optimize kernel parameters based on genetic algorithm(GA) isproposed.Taking Gauss function for instance, the influence of parameter selection uponkernel performance is analyzed. Then the method to optimize kernel parameters based on genetic algorithm(GA) is proposed because GA is stronge in global optimization. The results of experiment show that the kernel parameters educed by the method improve the kernel performance, and the method consumes much less time than cross-validation. By using the method based on GA ,the performance of three kinds of kernel function are compared. Taking error rate as fitness, parameter optimization experiment of kernel Gauss, kernel cubic polynomial, and kernel quartic polynomial are conducted. By analyzing the obtained data, the results show that: kernel Gauss is robust in classification, but it has some drawbacks such as more support vectors, large computation, slow detection speed. Whereas kernel polynomial is on the contrary.(2) Applying genetic algorithm based kernel parameter optimization method to face detection, a hierarchical SVM classifier is provided and realized.Since kernel Gauss and kernel polynomial have their own advantages and disadvantages, it is good to combine them. Taking the amount of support vectors as fitness to optimize kernel quartic polynomial’s parameter, then fast processing speed can be achieved. Taking error rate as fitness to optimize kernel Gauss’s parameter, then accurate recognition can be achieved. First, applying kernel polynomial SVM to input image, several candidate face windows are obtained. Then these candidate face windows are fused, and the output of the fusion are applied to kernel Gauss SVM. Finally, the ultimate results can be obtained. And a skin detector removing background beforehand is a plus. The author believes that, the method of combining both efficiency and accuracy can be used to construct a robust face detecting system.更多还原