Disjunctive Normal Parametric Level Set With Application to Image Segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Computational Geometry

Background:

• Level set methods are standard for image segmentation due to shape representation and topological change handling.
• Existing level set methods suffer from irregularities, initialization sensitivity, lack of locality, and high computational costs, especially for multi-object segmentation.

Purpose of the Study:

• To propose a novel parametric level set method, Disjunctive Normal Level Set (DNLS), for improved image segmentation.
• To address the limitations of conventional level set methods in terms of computational efficiency, regularity, and robustness.

Main Methods:

• Developed DNLS, a differentiable model using unions of polytopes derived from half-space intersections.
• Formulated the segmentation algorithm within a Bayesian framework, employing a variational approach for energy minimization.
• Applied DNLS to both two-phase and multiphase image segmentation scenarios.

Main Results:

• DNLS demonstrates significantly reduced computational time and memory requirements compared to conventional level sets.
• The method inherently maintains level set function regularity during evolution.
• DNLS exhibits enhanced suitability for multiphase, local region-based segmentations, and improved robustness to noise and initialization.

Conclusions:

• The proposed DNLS method offers a powerful and efficient alternative for image segmentation.
• DNLS provides a flexible framework adaptable to various appearance models and shape priors.
• Experimental results validate the effectiveness and advantages of DNLS over existing level set techniques.