Suppressed Kernel Possibilistic C-Means Clustering Based on Morphological Reconstruction for Image Segmentation

Spatial information is always introduced into the c-means clustering algorithms based on Euclidean distance to improve the segmentation accuracy of color images. However, most of these improved clustering algorithms failed to segment small targets in noisy images. Moreover, the introduction of spatial information usually leads to higher computational complexity due to the repeated calculation of the distances between the cluster centers and their local neighborhoods. The kernel-based possibilistic c-means algorithm (KPCM) exhibits strong anti-noise performance for data containing noise and outliers. However, its performance in noisy image segmentation is hindered by insufficient spatial information. To tackle the above problems, we propose a suppressed kernel possibilistic clustering segmentation (MRSKPCM) algorithm based on morphological reconstruction (MR). Firstly, local spatial information obtained by morphological reconstruction is integrated into the kernel-based possibilistic c-means clustering algorithm to ensure the advantages of anti-noise robustness and detail preservation. Secondly, the suppressed competitive learning mechanism is introduced to overcome the center overlapping problem by suppressing the possibilistic memberships of partial objects. Finally, the membership filtering technology is introduced to further improve the anti-noise robustness. Compared with the several state-of-the-art clustering algorithms, the MRSKPCM algorithm proposed in this paper not only segments successfully the small targets of imbalanced images but also enhances the algorithm efficiency. Additionally, the incorporation of membership filtering enhances the segmentation accuracy of images with noise injection and complex backgrounds.