Unsupervised texture segmentation based on FCM

作者：Jiang Xiaoyue; Zhao Rongchun; Jiang Zetao

来源：Journal of Computer Research and Development, 2005, 42(5): 862-867.

DOI：10.1360/crad20050522

As the cluster number of texture in an image is always unknown, the unsupervised classification is more valuable than the supervised classification. Based on the concept of a good cluster which should have the minimum intra-cluster distance and the maximum inter-clusters distance, the ratio of intra-cluster to inter-cluster distance is applied as the validity function. However, the increase of initial cluster number will influence the sum of cluster diameters and the inter-cluster separation distance. Therefore the maximum cluster diameter and minimum inter-cluster separation distance are provided instead, which is influenced by the initial cluster number more slightly and shows the essential of the cluster structure. Due to the relationship of FCM convergent speed with the initial cluster number, the convergent speed is introduced as the penalization factor to the validity function and a new validity function nRII is proposed. Compared with other validity functions, the nRII validity function can effectively prevent the over-clustering problem and give out a more exact estimation of the cluster number.

• 出版日期2005