支持向量机算法研究及其在目标检测上的应用
发布时间：2018-08-22 18:55
【摘要】：机器学习(Machine Learning,ML)是计算机科学的一项重要分支。Arthur Samuel将其定义为无需精确编程而能够具有学习能力的机器,也就是机器学习在一定程度上赋予了计算机一种"思考"能力。机器学习与计算统计学、数学优化和数据挖掘有着密切的关系,但也有一定的差异。根据学习过程中是否有反馈信息,机器学习可以分为监督式学习、无监督式学习和增强学习三大类。支持向量机(SVM)是一种监督式学习算法,可以用于解决分类、回归问题等。支持向量机是通过在特征空间构建超平面对分类或回归问题进行处理,通过核技巧将线性模型扩展到非线性情况。支持向量机是机器学习领域最优秀的算法之一,可以用于解决实际应用中的许多问题。支持向量机在文本和超文本分类、图像分类、手写识别、生物识别等多个方面取得广泛的应用。针对支持向量机核参数和其他相关参数的设置通常是基于经验的,而参数的选择常常关系到模型最终的性能,本文提出使用粒子群算法和人工蜂群算法对支持向量机参数的选择进行优化。相比其他SVM参数选择方法,由于群智能优化算法具有不要求参数连续、能够跳出局部极值等优点,因而基于群智能优化算法参数选择的SVM模型能够表现出更好的泛化性能。十几年前,使用机器对图像或视频中的目标识别和目标检索还是一件不可能实现的任务。但是随着近年来互联网的普及,越来越多的图像出现在互联网上,海量的图像数据使得人工图像处理和识别变得越来越不可能实现。研究人员一直致力于计算机视觉技术的研究,使得机器能够代替人工完成对图像的识别、分类、检索等任务。支持向量机算法在处理图像识别、分类、检索等多种计算机视觉方面的任务上表现出优异的性能。E-SVM(Exemplar-SVM)是近期提出的一种使用单一正样本与一个负样本集训练出的线性SVM模型。该算法已经在目标检测、基于内容的图像检索(Content Based Image Retrieval,CBIR)等领域取得了很好的应用。该算法针对每一个正样本训练一个相应的线性SVM分类器,最终得到一个单样本线性SVM模型的集合。在PASCAL VOC 2007目标识别数据集的测试表明,E-SVM方法能够取得与当前最优的目标检测算法LDPM相匹敌的识别率。由于E-SVM模型是对每一个样本进行训练最后得到多个特异性较强的检测器,本文提出使用K-均值聚类的方法对E-SVM检测器进行处理,得到一组具有样本平均特征的检测器,这组检测器由于融合了相应目标的多个特征,使得检测器具有更好的泛化性能;并且聚类后的检测器数目大大降低了,在进行目标检测的时候,能够降低识别时间,提升检测效率。
[Abstract]:Machine learning (ML) is an important branch of computer science. Arthur Samuel defines it as a machine capable of learning without precise programming, that is, machine learning gives the computer a "thinking" ability to a certain extent. Machine learning is closely related to computational statistics, mathematical optimization and data mining, but there are some differences. According to whether there is feedback in the learning process, machine learning can be divided into three categories: supervised learning, unsupervised learning and reinforcement learning. Support Vector Machine (SVM) is a supervised learning algorithm, which can be used to solve classification and regression problems. Support vector machine (SVM) is used to deal with the problem of classification or regression by constructing a superplane in the feature space, and the linear model is extended to nonlinear cases by kernel techniques. Support vector machine (SVM) is one of the best algorithms in the field of machine learning and can be used to solve many problems in practical applications. Support vector machine (SVM) has been widely used in text and hypertext classification, image classification, handwriting recognition, biometric recognition and so on. The kernel parameters of support vector machines and other related parameters are usually based on experience, and the selection of parameters is often related to the final performance of the model. In this paper, particle swarm optimization and artificial bee swarm algorithm are used to optimize the parameters of support vector machine. Compared with other SVM parameter selection methods, the swarm intelligence optimization algorithm has the advantages of not requiring continuous parameters and being able to jump out of the local extremum, so the SVM model based on the parameter selection of the swarm intelligence optimization algorithm can show better generalization performance. More than a decade ago, using machines to identify and retrieve targets from images or videos was an impossible task. However, with the popularity of the Internet in recent years, more and more images appear on the Internet. The massive image data make artificial image processing and recognition more and more impossible. Researchers have been devoting themselves to the research of computer vision technology, which enables machines to complete the tasks of image recognition, classification, retrieval instead of manual. The support vector machine (SVM) algorithm shows excellent performance in many computer vision tasks, such as image recognition, classification, retrieval and so on. E-SVM (Exemplar-SVM) is a recently proposed linear SVM model trained using a single positive sample and a negative sample set. The algorithm has been applied to target detection and content-based image retrieval (Content Based Image Retrieval CBIR). The algorithm trains a corresponding linear SVM classifier for each positive sample and finally obtains a set of single-sample linear SVM models. The test results on PASCAL VOC 2007 target recognition data set show that the proposed method can achieve recognition rate comparable to the current optimal target detection algorithm (LDPM). Because the E-SVM model is to train each sample to obtain several detectors with strong specificity, this paper presents a K- mean clustering method to process the E-SVM detector and obtain a set of detectors with average sample characteristics. Due to the fusion of multiple features of the corresponding target, these detectors have better generalization performance, and the number of detectors after clustering is greatly reduced, and the recognition time can be reduced when the target is detected. Improve detection efficiency.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18
[Abstract]:Machine learning (ML) is an important branch of computer science. Arthur Samuel defines it as a machine capable of learning without precise programming, that is, machine learning gives the computer a "thinking" ability to a certain extent. Machine learning is closely related to computational statistics, mathematical optimization and data mining, but there are some differences. According to whether there is feedback in the learning process, machine learning can be divided into three categories: supervised learning, unsupervised learning and reinforcement learning. Support Vector Machine (SVM) is a supervised learning algorithm, which can be used to solve classification and regression problems. Support vector machine (SVM) is used to deal with the problem of classification or regression by constructing a superplane in the feature space, and the linear model is extended to nonlinear cases by kernel techniques. Support vector machine (SVM) is one of the best algorithms in the field of machine learning and can be used to solve many problems in practical applications. Support vector machine (SVM) has been widely used in text and hypertext classification, image classification, handwriting recognition, biometric recognition and so on. The kernel parameters of support vector machines and other related parameters are usually based on experience, and the selection of parameters is often related to the final performance of the model. In this paper, particle swarm optimization and artificial bee swarm algorithm are used to optimize the parameters of support vector machine. Compared with other SVM parameter selection methods, the swarm intelligence optimization algorithm has the advantages of not requiring continuous parameters and being able to jump out of the local extremum, so the SVM model based on the parameter selection of the swarm intelligence optimization algorithm can show better generalization performance. More than a decade ago, using machines to identify and retrieve targets from images or videos was an impossible task. However, with the popularity of the Internet in recent years, more and more images appear on the Internet. The massive image data make artificial image processing and recognition more and more impossible. Researchers have been devoting themselves to the research of computer vision technology, which enables machines to complete the tasks of image recognition, classification, retrieval instead of manual. The support vector machine (SVM) algorithm shows excellent performance in many computer vision tasks, such as image recognition, classification, retrieval and so on. E-SVM (Exemplar-SVM) is a recently proposed linear SVM model trained using a single positive sample and a negative sample set. The algorithm has been applied to target detection and content-based image retrieval (Content Based Image Retrieval CBIR). The algorithm trains a corresponding linear SVM classifier for each positive sample and finally obtains a set of single-sample linear SVM models. The test results on PASCAL VOC 2007 target recognition data set show that the proposed method can achieve recognition rate comparable to the current optimal target detection algorithm (LDPM). Because the E-SVM model is to train each sample to obtain several detectors with strong specificity, this paper presents a K- mean clustering method to process the E-SVM detector and obtain a set of detectors with average sample characteristics. Due to the fusion of multiple features of the corresponding target, these detectors have better generalization performance, and the number of detectors after clustering is greatly reduced, and the recognition time can be reduced when the target is detected. Improve detection efficiency.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18
【参考文献】
相关期刊论文 前4条
1 陈健飞;蒋刚;杨剑锋;;改进ABC-SVM的参数优化及应用[J];机械设计与制造;2016年01期
2 李t熋，
本文编号：2198007
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2198007.html
