Abstract

Image segmentation remains a fundamental challenge in computer vision, where multilevel thresholding becomes computationally demanding as the number of thresholds increases and conventional optimisation methods often converge prematurely. This study integrates chaotic maps into the Particle Swarm Optimisation (PSO) algorithm to enhance population diversity, prevent premature convergence, and improve segmentation quality. The aim is to evaluate three chaos-enhanced PSO (CE-PSO) variants, Logistic+PSO, Sinusoidal+PSO and Gaussian+PSO, for image segmentation using multilevel thresholding on the Berkeley Segmentation Dataset (BSDS500). Chaotic sequences were used in place of the pseudo-random number generators employed for population initialisation, parameter adaptation, and position updates. The models were implemented in Python and evaluated on the preprocessed BSDS500 dataset using PSNR, SSIM, FSIM, Dice Coefficient, and Jaccard Index, averaged over 30 independent runs per algorithm. Sinusoidal+PSO achieved the best results, with a PSNR of 27.89 dB, SSIM of 0.9123, FSIM of 0.8967, Dice Coefficient of 0.8876, and Jaccard Index of 0.8234, followed by Gaussian+PSO (27.12 dB, 0.9034, 0.8856, 0.8765, 0.8067) and Logistic+PSO (26.78 dB, 0.8912, 0.8745, 0.8654, 0.7923); all three variants outperformed Standard PSO on every metric. Sinusoidal+PSO also converged fastest, reaching its solution in an average of 41 iterations, a 29.3% improvement over Standard PSO. The study establishes that Sinusoidal+PSO offers the best accuracy-efficiency trade-off, Gaussian+PSO offers strong mixing properties for problems requiring diversified search, and Logistic+PSO offers a computationally simple reference implementation.