关键词: 图像检索, 支持向量机, 属性约简.

Abstract:

Based on the visual semantics of images,a novel space-transforming model is designed,and a new semantics description method is proposed.In the investigation,first,the method of Normalized Cut（NCut） is used to segment each image into several regions,and such visual features as color,texture and shape,etc.of each region are extracted.Next,all the visual features in the training set are clustered by using the K-Means method,and each cluster center is regarded as a ＂visual semantic＂ to construct a projection space.Then,a nonlinear function is defined to map each image into a point in the projection space.Thus,all the projection features of the image are obtained.Moreover,in order to improve the training efficiency and performance of classifiers,projection features are reduced in attributes via the rough set（RS） method,and are trained and classified by the support vector machine.The results are finally compared with those based on the Corel image set.It is found that the proposed method is robust to cluster number,and is superior to the other methods in terms of retrieval performance.

Key words: image retrieval, support vector machine, attribute reduction