Title: A novel superpixel segmentation method for improved image segmentation using soft probabilities and local pixel variation
Authors: Ke Liu
Addresses: College of Information Science and Technology, Jinan University, Guangzhou, Guangdong, 511443, China
Abstract: Superpixel segmentation is a critical technique in computer vision, widely used in tasks such as image segmentation and object recognition. However, traditional methods struggle with complex and irregular boundary regions, often resulting in inaccurate boundary delineation and irregular superpixel shapes. To address these challenges, this paper proposes a novel superpixel segmentation method that introduces soft probabilities and leverages local pixel variation, combined with a weight function derived from edge detection to enhance global contextual information. Soft probabilities allow smoother transitions between pixel regions, while local pixel variation helps capture fine-grained details in boundary regions. Evaluations on the improved BSDS500 dataset show a 4% region similarity improvement, 8% information entropy reduction, and moderate accuracy gains over the best current method. The results highlight the method's effectiveness in managing complex boundary information while maintaining high segmentation accuracy, representing a significant advancement in superpixel segmentation.
Keywords: superpixel segmentation; edge detection; soft probabilities; image boundary segmentation; local pixel variation.
DOI: 10.1504/IJCSM.2025.147104
International Journal of Computing Science and Mathematics, 2025 Vol.21 No.2, pp.109 - 123
Accepted: 27 Feb 2025
Published online: 10 Jul 2025 *