image classification with ant colony based support vector machine
- 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。
Image Classification with Ant Colony Based Support Vector Machine
ZHAO Baoyong1,2, QI Yingjian3
1. School of Automation & Electrical Engineering, University of Science and Technology Beijing, Beijing 100083, P. R. China
2. Key Laboratory of Advanced Control of Iron and Steel Process, University of Science and Technology Beijing,
Beijing 100083, P.R. China.
E-mail: zbyfirst@
3. Science School, Communication University of China, Beijing 100024, P.R. China
E-mail: qiyingjian@
Abstract: Natural image classification is an important task. SIFT descriptors and bag-of-visterms (BOV) method have achieved very good results based on local image representation. Many studies use the support vector machine to classify and identify the image category after finished representation of the image. However, due to support vector machine (SVM) its own characteristics, it shows inflexible and less slow convergence rate. The selection of parameters influenced the results for the algorithm seriously. Therefore, this paper try to improve the image recognition performance by support vector machine algorithm based on ant colony algorithm. The method adopt dense SIFT descriptors and BOV method to obtain the image representation. In recognition step, we use the support vector machine as a classifier but ant colony optimization method is used to selects kernel function parameter and soft margin constant C penalty parameter. Experiment results show that this solution determined the parameter automatically without trial and error and improved performance on natural image classification tasks. Key Words: Scale Invariance Feature Transform, Bag-of-Visterms, Colony Algorithm, Support Vector Machine
1Introduction
Classification of natural image automatically is one of the significant research fields in computer vision. Effective image description is the base of the correct image classification. And then how to discover the objects and use the objects distribution to perform scene classification is regarded as in an unsupervised manner. Currently, the method based on semantic analysis represents the image mainly through the use of intermediate semantic (or Topic) [1, 2]. Each image can be expressed by a number of Topics which based on the probability distribution of the so called visual words (visual visterms).
Recently, Bag-of-visterms (BOV) becomes an important image content representation method [3]. Imitation the idea of text classification, the semantic information is exhibition of the words distribution. The image can be seen as composed by the visual words document, and then the visual words used to classify the image. However, BOV's problem is that the image is difficult to define specific meaningful visual words, because it needs to transform the image characteristics into the visual terms effectively. A lot of work in this area has done: Some used the method that found the dense sampling local area and extracted the feature vector in the region to express the image content, thus greatly improving the performance of the visual vocabulary [4, 5]. Some detected the interest image region, and then extracted the SIFT descriptor in the region as a feature vector [6, 7]. Clustering and classification of vocabulary commonly used k-means clustering method, but improved algorithm combined with other optimized cluster method to achieve improvements on the classifier has continued [8-10].
Support Vector Machine (SVM) is a new learning machine developed on the basis of structural risk *This work is supported by the project from Beijing Key Discipline Development Program (No. XK100080537) minimization principle [11]. It aimed at finding compromise between complexity of model and the best learning ability in the based of limited sample information for good generalization ability. Support vector machine is based on solid statistical learning theory and have theoretical completeness, but there are still some problems on the application, one of the typical problems is the choice of model parameters. There is no uniform model selection standards and theory. Penalty factor C and kernel function parameters have a major impact on the classification performance in specific applications [12, 13]. Ant colony algorithm is a new heuristic evolutionary algorithm. It has a strong ability to find better solutions, better robustness, information positive feedback, parallel distributed computing and easy to combine with other heuristic methods [14]. It is a good idea to use ant colony algorithm in support vector machine which provides powerful method for SVM parameter selection problem.
This paper gives a new method that uses the bag of word method in creating an image described, the neighborhood features of the image is described by using 128-dimensional sift descriptor, and then using two-level vector quantization to build vocabulary, while getting semantic description of the image by the pyramid dense sampling. In the image recognition stage, SVM classifier [15] is used in recognition and the parameters of SVM is optimized by using ant colony algorithm [16], and test it in the image database. Experimental results show that the method can automatically determine the SVM parameters and can better improve the performance of the natural image classification.
Proceedings of the 30th Chinese Control Conference July 22-24, 2011, Yantai, China