Q-Seg: Quantum Annealing-Based Unsupervised Image Segmentation.

Area of Science:

• Quantum Computing
• Computer Vision
• Image Segmentation

Background:

• Unsupervised image segmentation is crucial for analyzing complex visual data.
• Existing quantum and classical methods face challenges in scalability and performance.
• Earth observation data often suffers from noisy and unreliable annotations.

Purpose of the Study:

• Introduce Q-Seg, a novel unsupervised image segmentation method.
• Leverage quantum annealing for efficient graph-cut optimization.
• Evaluate Q-Seg's performance against classical methods and advanced techniques.

Main Methods:

• Formulate pixelwise image segmentation as a graph-cut optimization problem.
• Utilize the D-Wave Advantage quantum annealing architecture.
• Assimilate spectral and spatial image information.

Main Results:

• Q-Seg demonstrates superior scalability compared to existing quantum approaches.
• Outperforms state-of-the-art classical methods in empirical evaluations.
• Shows better runtime performance than the classical optimizer Gurobi on synthetic datasets.
• Successfully applied to earth observation image segmentation tasks.

Conclusions:

• Q-Seg offers a promising unsupervised image segmentation solution using available quantum hardware.
• Provides a reliable alternative for real-world applications, especially with limited labeled data.
• Competes effectively with advanced techniques like Segment Anything, particularly in noisy environments.