基于保局映射的图像纹理聚类

Image texture clustering based on locality preserving projection

为了解决图像检索中的聚类问题,提出一种改进的图像纹理聚类算法.在纹理特征提取阶段,采用双树复小波对图像进行分解,然后对每个高频段提取直方图签名作为纹理特征;在聚类阶段,根据数据分布的密度来动态地计算数据点的邻接矩阵,再采用保局映射进行降维,对降维后的数据进行k-means聚类.通过采用直方图签名的方式能有效地表示图像纹理在各个方向上特征信息,同时根据数据密度构建的邻接矩阵,能够和保局映射一起更有效地发掘数据之间的局部相关性.实验表明:相对于传统方法,该算法具有更高的聚类正确性.

An improved image texture clustering method was proposed to solve the clustering problem in image retrieval. In feature extraction stage, dual-tree complex wavelet transform（DT-CWT）is applied to decompose image into tens of sub-bands. For those high frequency sub-bands, histogram signatures are generated as one of the texture features. In clustering stage, the distances between the data points are computed adaptively according to the data distrihution density. I.ocality preserving projection is then employed on the distances to reduce the dimensionality of the data space, k-means is used to cluster the data in the lower dimensionality space. Histogram signature can represent image texture well in multiple directions of the DT CWT decomposition. Moreover, the distance matrix built on data density can detect dataset locality effectively in combination with locality preserving projection. The experimental results show that the proposed method outperforms the traditional methods.