融合复小波特征和局部二值模式的纹理聚类

为了解决图像检索中聚类问题,对图像作双树旋转复小波变换,再对变换后的系数通过广义高斯模型建模后,计算Kullback-Leibler距离;对图像采用局部二值模式,计算图像之间的对数似然距离.将这2种距离采用事先计算得到的加权因子进行融合得到新的距离.在此距离的基础上构建邻接矩阵,对邻接矩阵采用谱聚类的方法进行聚类运算.实验证明,由于双树旋转复小波变换和局部二值模式之间存在互补性,在聚类过程中将2种特征距离结合起来,能够有效地提高聚类的正确性.

Texture clustering combining complex wavelet features and local binary pattern

A dual tree rotated complex wavelet (DT RCW) transform was applied on images to solve the clustering problem in image retrieval. Kullback Leibler distances were computed with the generalized Gaussian density model of each high frequency band. A local binary pattern was employed on each image and the log likelihood distances were computed. The two distances were combined with the pre computed weight to produce new distance, on which an adjacent matrix was built. Then spectral clustering was performed with the generated adjacent matrix. Because the local binary pattern can be considered as the complementary feature of rotated complex wavelet, the experimental results show that combining the rotated complex wavelet and the local binary pattern can effectively improve the clustering performance.