联系
Knight's Tale » 我的作品

基于支持向量机语义分类的两种图像检索方法

2010-08-02 19:43

This paper has been accepted by JOURNAL OF XIAMEN UNIVERSITY(NATURAL SCIENCE), 2010, No.4 and the 6th Conference on Intelligent CAD and Digital Entertainment.

基于支持向量机语义分类的两种图像检索方法

廖绮绮,李翠华(厦门大学计算机科学系,福建省厦门市,361005)

摘  要 为了更好的解决基于内容的图像检索问题,提出了两种基于语义的图像检索方法。第一种是基于SVM语义分类的图像检索方法。该方法首先提取训练图像库的底层特征信息,然后利用支持向量机对所提取的特征进行训练,构造多分类器。在此基础上,利用分类器对测试图像自动分类,得到图像属于各个类别的概率,实现图像检索。第二种是利用图像自动标注方法进行检索。在基于语义的图像自动标注中,先对训练集进行人工标注,对测试图像利用SVM分类器进行分类,并找到与该图像最相似的N张构成图像集,对该图像集的标注进行统计,找到关键词,从而提供概念化的图像标注以用于检索。通过在标准图像检索库和自建图像库上的实验结果表明,本文提出的两种基于语义的图像检索方法是高效的。

关键词 图像检索;语义特征;SVM支持向量机;分类器

中图分类号:TP391. 41             文献标识码:A

Two Image Retrieval Methods Based on Support Vector Machines Semantic Classification

Liao Qi-Qi,Li Cui-hua(Department of Computer Science and Technology, Xiamen University,Xiamen,361005)

Abstract: In order to solve the problem of content based image retrieval(CBIR), two novel methods of image retrieval based on semantic are proposed. Firstly, we use the method of image retrieval which is based on the SVM semantic classification. In this method, we will use Support Vector Machines (SVM) Statistical Learning Theory tools to train the visual image features in order to construct Multi-class Classifier. Thus the test images can be automatically classified by using this classifier, and we will get the probability of the images belong to the every class easily. Then we use this probability to compute the similarity between images. The second method is the image retrieval based on automatic image annotation. Based on the SVM semantic classification method, the image database which is noted will be used as the training image set. Then we use SVM classifier to classify the test images and find the N-nearest similar images in the image library. Then we estimate the probability of the key words from those images and the automatic image annotation will be accomplished. And this image annotation will be used to image retrieval. Experiments conducted on standard dataset and realistic dataset demonstrate the effectiveness and efficiency of the proposed approach for image retrieval.

Key words: image retrieval; semantic feature; SVM support vector machines; classifier