【作者】 杨小艳；

【导师】 葛洪伟；

【作者基本信息】 江南大学 ， 计算机软件与理论， 2007， 硕士

【摘要】 支持向量机(SVM)是建立在统计学习理论(SLT)的VC维理论和结构风险最小原理基础上的一种新的通用学习方法,它根据有限的样本信息在模型的复杂性和学习能力之间寻求最佳折衷,以期获得最好的推广能力。目前,在模式识别方面,SVM算法在精度上已经超过传统的学习算法或与之不相上下。核函数是SVM方法中少数几个能够调整的参数之一,由于核函数的形式及其参数的确定决定了分类器类型和复杂程度,它显然应该作为控制分类器性能的手段。本文针对实际问题,即针对一款基于支持向量机的羽绒种类自动检测系统,重点对核函数及其参数的选择进行了探讨和研究,看能否进一步提高该系统的识别率。基于此,本文的主要工作如下:1、在对图像处理、SVM理论等知识进行了探讨的基础上,对内核及其参数的选择进行了重点研究,并提出利用优化算法选择最佳内核参数来提高系统的性能。通过对优化算法及内核优化思想的的研究,来确定最佳内核及其参数,从而在对羽绒图像中的菱节进行识别时,识别率能有显著提高,之后对识别出的菱节进行配对、计算节距来最终确定羽绒的种类为鹅绒、鸭绒或非鹅鸭绒时,使整体识别率得到提高。2、对构建新内核进行了研究,并将之与内核优化思想相结合进行了探讨。通过对多个内核的测试比较,使羽绒种类自动检测系统中所选内核及其参数为最佳,并完善了羽绒种类自动检测系统的软件设计。在Windows 2000下,以VC++ 6.0为开发平台最终开发出了界面友好、操作简单、识别结果比较满意的羽绒种类自动检测系统软件。基于SVM的羽绒种类自动检测系统克服了人工检测的诸多弊端,减少了人工检测时的可能出错,是羽绒种类检测方式的一种尝试和突破;本系统的研究开发,大幅提高了羽绒鉴定的手段与水平,具有广阔的前景和推广价值,并且填补了国内外在此领域的空白。更多还原

【Abstract】 Supprot Vector Machine(SVM) is a new general learning method developed in recent years on the goundation of Structural Risk Minimization principle and VC Dimension of Statistical Learning Theory and by searching the best trade-off base on the limited samples information between pattern complexity and learning ability. Currently, SVM algorithm is more precise than or equal to traditional algorithm in pattern recognition .Kernel function is one of the tunable parameters in SVM classifier. Since form and parameters of kernel function decide the type and the complexity of classifier,it should act as a means of controling the performance of classifer. In this paper, we address the selective problem of kernel function and its parameters in SVM modeling. Our work is proceeding in down category recognition and address improving the recognition rate of system,which run as follows:1.On the base of researching theories of SVM and the methods of image processing,we address the study of SVM kernel and its parameters, and SVM kernel optimization method is proposed to improve the performance of system. In down recognition,we decide the best kernel and its parameter values by the method and improve the recogniton rate of the triangle nodes of down with the SVM model, so the triangle nodes which have been recognized can be matched each other and the distance between the matched triangle nodes is calculated. in the end, the down category which is duck’s down or goose down,or other one is recognized and integral recognization rate can be improved.2.We make a study of the problem about constructing a hybrid kernel, then combine kernel construction method and kernel optimization method and research together. After test and compare the different types of kernels, we select the optimal kernel and its parameters for down category recognition system, and the software of down category automatic recognition system is designed and perfected. In the end, we develop a system with friendly interface and simple operation with VC++ 6.0 in Windows2000.In the paper, the down category automatic recognition base on SVM conquers some disadvantages and inconveniences of manual identification, and decreases the faults of manual identification. Besides, the system can improve the level of down category recognition and our study has a wide range of further use and practicality. Even it fills up the blank in the domain at home and abroad.更多还原