Semi-supervised possibilistic Gustafson-Kessel clustering algorithm based on SLIC and MR

The possibilistic c-means clustering (PCM) modifies the condition that the sum of the fuzzy memberships of each sample point is one, which improves the robustness of the fuzzy c-means algorithm (FCM) to noise and outliers. The semi-supervised possibilistic c-means clustering algorithm (SS-PCM) improves the accuracy of datasets with small targets to a certain extent by introducing partial prior information. The SS-PCM algorithm cannot handle datasets with strong feature correlation and imbalanced sizes. We propose a semi-supervised possibilistic Gustafson-Kessel clustering algorithm (SM-SS-PGK) based on simple linear iterative clustering (SLIC) and morphological operation (MR). In addition, the SM-SS-PGK algorithm improves the robustness of the anti-strong mixed noise and retains image details. Firstly, the Mahalanobis distance is introduced into the objective function (SS-PCM), which can effectively deal with strong feature correlation datasets. Secondly, to enhance the utilization of supervised information, supervised centers are introduced to guide the center-iteration of small clusters and reduce the initialization sensitivity. In addition, to improve the noise robustness of the algorithm, we introduce SLIC superpixels and morphological reconstruction (MR) to preserve image detail information. Finally, color image segmentation experiments show that the SM-SS-PGK method achieves better segmentation results than several clustering algorithms.