基于小波多尺度特征的图像聚类检索

Wavelet Multi—scale Features Clustering Based Image Retrieval

描述了一种图像数据库中基于小波多尺度特征内容的聚类检索方法。该方法对图像数据库中的图像进行小波多尺度分解并提取每一频段的矩和最低频段的小波系数分别作为其纹理特征和颜色特征。为提高检索效率,在图像被插入到图像数据库时对其进行基于多尺度矩的K均值聚类。检索时,将查询图像与聚类各簇的质心进行比较确定其相似簇,加上颜色特征计算查询图像与相似簇中各图像的相似性距离。实验证明：该方法由于综合考虑图像的纹理和颜色特征信息,因而具有较高的查准率和查到率,而聚类算法的应用使其有较高的检索速度。

A wavelet multi-scale features clustering based Image retrieval approach is proposed in this paper. This approach applies wavelet multi-scale decomposition to each image in image database and then extracts moments of every sub-band and lowest frequency sub-band wavelet dominant coefficients histogram as texture features and color features. Then a clustering techniques is developed to reduce the query time. Retrieval procedure is consist oftwo steps： ①determine similar cluster by compare texture distance between query image and centroid of every cluster. ②determine similar image by compute distance between query image and every image in similar cluster. The prototype system test demonstrate that this approach has high retrieval performance.