A cutset-type kernel possibilistic fuzzy c-means method for robust image segmentation

Robust image segmentation remains difficult for lightweight unsupervised clustering when nonlinear class overlap and sparse noise contamination occur together.

This study proposes and evaluates a cutset-type kernel possibilistic fuzzy c-means method (C-KPFCM), combining Gaussian-kernel distance modeling with cutset-based correction of possibilistic typicalities. The kernel term changes the distance geometry for nonlinear structure, while the cutset operation suppresses non-winning typical values during center estimation.

On the three reconstructed complex-background cases, C-KPFCM obtains the best average result among the tested classical fuzzy clustering methods, but the gain over the closest cutset-only baseline is small and not statistically reliable under an exact sign test. A 90-instance stress test further shows that C-KPFCM is clearly better than K-means, PFCM, and KPFCM, but not generally better than C-PFCM under broader disturbance mixtures. Additional noise, parameter-sensitivity, repeated-run, runtime, and real-image pilot analyses support a restrained conclusion.

Cutset correction is the empirically dominant robustness mechanism, while kernelization is a conditional complement whose cost is justified only when nonlinear geometry is expected to matter. Code and reproducibility scripts are available at: jtwang-AI/Image-Segmentation-C-KPFCM.