核空间聚类在图像纹理分类中的简化算法

Simplified method of kernel fuzzy c-means clustering for image texture classification

模糊c均值聚类已广泛应用于模糊模式识别领域,但对于线性不可分数据并不适用.在核方法中通过将输入数据经过非线性映射投影到高维特征空间来解决非线性分类的问题.将传统的模糊c均值聚类算法应用于核空间中,对线性不可分的样本进行了核空间聚类的分类实验,得到了正确的分类结果.由于图像分类中分类样本(对应图像像素)数目庞大,造成了核空间聚类算法中特征距离的计算量过大.因此,在核空间聚类的基础上,提出了对图像先进行过分割,再对过分割的图像块进行核空间聚类的方法,大大降低了高维空间特征距离计算的运算成本,并取得了良好的分类效果.

The fuzzy c-means clustering algorithm is a widely applied method for acquiring fuzzy pattern from data, but it is not suitable for the clustering of linear inseparable data. In mercer kernel method, the problem of nonlinear separability of classes can be tricked by projecting the input data to a higher dimensional feature space in a nonlinear manner. So the fuzzy c-means clustering method was used in the mercer kernel space. The classification experiment illustrated that the kernel fuzzy c-means clustering （KFCM） algorithm was suitable for the clustering of linear inseparable data. When KFCM clustering was used in image segmentation, the large number of classification samples always caused the computational burden. The image classification procedure was divided into two steps： firstly, the image was over-segmented into large numbers of small regions according to the input features; secondly, they were classified with KFCM. The computational burden was reduced by the decrease of classification samples, while the classification result was almost as good as KFCM＇s.