Enhancing decision support for pattern classification via fuzzy entropy based fuzzy C-Means clustering

Digital image segmentation is widely applied to split a visual representation into distinct clusters for pattern classification and target detection. In contrast to the C-Means clustering with hard clusters, the generalized fuzzy C-Means clustering provides soft clusters such that each image pixel belongs to multiple clusters with fuzzy degrees of belonging. The fuzzy C-Means algorithms involve the integration of the intensity, color, texture and position to partition the feature space into multiple regions, while the boundary information is seldom taken into account. Thus extensive research should be conducted for improvement. The proposed fuzzy entropy based fuzzy C-Means clustering, on the other hand, is able to locate vague boundaries that any crisp clustering can hardly reach. Adoption of the notion of the fuzzy entropy into fuzzy C-Means clustering enables the boundary information to be recovered. Optimization can be achieved by comprising both the contour information and the well-known classical region based C-Means segmentation. To quantify the potential benefit of the proposed approach, a comparative study is made on fuzzy mutual information computed from two sets of patterns that have been generated before and after the fuzzy entropy is included. The simulation outcomes indicate the merits of the novel scheme.