提出一种基于多子空间KL变换的纹理图像自监督分割方法。该方法将非监督聚类转变为有典型特征样本指导的自监督分类，解决误分类率高的问题。采用多子空间方法对样本进行特征选择，克服假设所有纹理特征都属于单个高斯分布的局限性。首先，对待分割图像进行多尺度、多方向的Gabor变换，使用模糊C均值方法从变换结果中提取具有典型性的样本作为训练样本；然后，使用训练样本为每一个类别生成一个单独的初始子空间；最后，采用多子空间KL变换，对其余样本在迭代过程中进行类别划分。实验结果证明，本文方法能够减少误分类率，改善分割效果。

This paper presents a texture segmentation algorithm based on self-supervised classification and multispace KL transform. It turns unsupervised clustering into self-supervlsed classification to decrease the ratio of misclassificatlon. Our algorithm adopts a multispace method for feature selection to avoid the limitations introduced by supposing that all samples obey a single Gauss distribution. Firstly multldirection and multiscale Gabor transforms are applied to target texture images ; then fuzzy C means clustering is acted on the results of above transforms to extract some typical training samples, which are requested to supervise later segmentation. Secondly a separate subspace for each class is initialized by training samples respectively. Lastly other samples are classified with multispace KL transforms through the iterative processes. Our algorithm is fully competent for various composite texture segmentations. And experimental results have proved that it can successfully reduce misclassification ratio in the same time improve the visual effects of texture segmentation.