A Continuous Random Walk Model With Explicit Coherence Regularization for Image Segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Computational Mathematics

Background:

• Random walk algorithms are effective for image segmentation and region of interest (ROI) extraction.
• Traditional random walk methods face challenges with large datasets due to computationally intensive linear systems and sensitivity to seed distribution.
• User interaction is often burdensome due to the need for precise seed placement.

Purpose of the Study:

• To develop a continuous random walk model with coherence regularization (CRWCR) to mitigate seed sensitivity in image segmentation.
• To create an efficient algorithm for solving the CRWCR model, addressing the computational burden of large sparse linear systems.
• To reduce user interaction required for accurate ROI extraction.

Main Methods:

• A novel continuous random walk model with explicit coherence regularization (CRWCR) is proposed.
• A two-stage algorithm is developed: 1D random walk sweeping for initialization, followed by the Peaceman-Rachford scheme for refinement.
• The algorithm is designed for efficiency, leveraging GPU computing capabilities.

Main Results:

• The CRWCR model demonstrates reduced sensitivity to seed distribution, simplifying user interaction.
• The proposed two-stage algorithm efficiently solves the CRWCR model, overcoming the challenge of large linear systems.
• Numerical experiments validate the model's effectiveness and the algorithm's computational efficiency, with typically 10 iterations for convergence.

Conclusions:

• The CRWCR model offers an improved approach to image segmentation by enhancing robustness to seed placement.
• The developed two-stage algorithm provides a computationally efficient solution for applying random walk segmentation to large images.
• This work reduces user burden and computational cost in extracting regions of interest.