基于旋转复小波变换的图像纹理谱聚类算法

Spectral Clustering Algorithm of Image Texture Based on Rotated Complex Wavelet Transform

纹理作为图像的重要信息,在图像检索中起着重要作用.本文提出一种基于图像纹理的聚类算法.首先采用双树复小波加旋转复小波分解图像,得到十二个方向的高频数据.然后对每个高频段提取直方图签名.通过把直方图签名作为纹理特征之一,来计算数据点之间的相似性,采用改进的谱聚类进行降维.最后,对降维后的数据进行K-means聚类.因为本文采用直方图签名的方式有效地表示了在双树和旋转复小波分解后各个方向上的特征信息,同时在谱聚类过程中,提出一种动态的方式,根据数据点密度来计算数据间的相似度,从而有效地发掘了数据之间的局部相关性.实验表明,本文算法能够较显著地提高聚类的正确性.

As an important feature, texture plays a critical role in image retrieval. A clustering method is proposed based on image texture. Rotated complex wavelet （RCW） and dual-tree complex wavelet transform （DT-CWT） are used to decompose image into high frequency coefficients in twelve directions. The histogram signatures can be computed from each high frequency sub-band. Combined with other features, those signatures are employed to compute the similarity between data points for the improved spectral clustering to reduce dimensionality. In the final step, k-means is applied on the dimensionality-reduced data to get the clustering result. The proposed histogram signature for RCW and DT-CWT decomposition can capture the high frequency information in each direction effectively. In addition, an adaptive approach is proposed to compute the similarity between data points in spectral clustering. The experimental results show the proposed method outperforms the traditional methods remarkably.