Linear boundary detection by cluster prototype centring based on fuzzy memberships

A method based on fuzzy clustering by the centring of prototypes on the basis of memberships is proposed for the detection of linear boundaries in digital images. The algorithm applies three rules, governing the updating of memberships, the updating of line gradients and the centring of prototypes, to generate solutions to linear clusters by the specifications of four basic parameters involving the minimum cluster size, the membership threshold of the cluster, the image scale factor and the fuzzy factor. Clusters in the image space are found by a cycle of clustering processes involving the iterative development of a single prototype, the removal of a valid cluster if one is found, or the removal of invalid data as noise points, and the updating of the data list. This cycle is repeated until all possible clusters are exhausted from the image space. Test results on a pentagon shaped edge-segmented object indicated the essential robustness and reliability of the algorithm to correctly detect five linear segments, even in the presence of considerable noise and blurring.