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

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Image segmentation is difficult with unsupervised clustering methods.
• Nonlinear class overlap and sparse noise contamination pose significant challenges.
• Existing methods struggle with robust segmentation in complex scenarios.

Purpose of the Study:

• To propose and evaluate a novel image segmentation method.
• To address limitations in unsupervised clustering for complex image data.
• To enhance robustness against nonlinear class overlap and noise.

Main Methods:

• Introduced a cutset-type kernel possibilistic fuzzy c-means (C-KPFCM) method.
• Utilized Gaussian-kernel distance modeling for nonlinear structures.
• Implemented cutset-based correction to refine typicality estimates.

Main Results:

• C-KPFCM achieved the best average results among tested fuzzy clustering methods on complex backgrounds.
• Demonstrated superior performance over K-means, PFCM, and KPFCM in stress tests.
• Found cutset correction to be the primary driver of robustness, with kernelization offering conditional benefits.

Conclusions:

• Cutset correction is a key mechanism for improving segmentation robustness.
• Kernelization complements cutset correction, particularly for nonlinear data geometries.
• The C-KPFCM method offers a valuable advancement for challenging image segmentation tasks.