A Fast and Robust Cluster Update Algorithm for Image Segmentation in Spin-Lattice Models Without Annealing. Visual...

Area of Science:

• Computer Vision
• Computational Physics
• Image Processing

Background:

• Image segmentation is crucial for identifying objects in spin-lattice models.
• Traditional local spin-update algorithms are slow and require careful annealing schedules.
• Existing cluster update methods can merge distinct objects.

Purpose of the Study:

• To develop a novel, efficient, and robust cluster update algorithm for spin-lattice models.
• To improve the speed and reliability of image segmentation.
• To enhance segmentation quality by incorporating luminance-dependent visual latencies.

Main Methods:

• Proposed the energy-based cluster update (ECU) algorithm.
• Calculated energy gain for flipping entire spin clusters.
• Introduced luminance-dependent visual latencies into the spin-lattice model.
• Validated convergence and performance against predecessors.

Main Results:

• The ECU algorithm significantly outperforms local update algorithms in speed and reliability.
• ECU is robust to noise and eliminates the need for annealing.
• Segmentation of real images is achieved in 1-5 seconds on a standard workstation.
• Incorporating visual latencies improved segmentation quality by 40%.

Conclusions:

• The ECU algorithm offers a substantial advancement in spin-lattice model-based image segmentation.
• The algorithm is efficient, robust, and adaptable to various image features.
• The integration of visual latencies further refines segmentation accuracy.