Image segmentation is a fundamental operation in computer vision, and multilevel thresholding is among the techniques most commonly used to perform it; however, as the number of thresholds increases, locating the best set of threshold values becomes computationally demanding, and conventional methods, together with the Particle Swarm Optimisation (PSO) algorithm ordinarily used to address this difficulty, frequently suffer from premature convergence and inadequate exploration of the search space. This study addressed these limitations by incorporating chaotic maps into PSO to sustain population diversity, discourage premature convergence, and improve segmentation accuracy. The aim was to evaluate three chaos-enhanced PSO (CE-PSO) variants, namely Tent+PSO, Sinusoidal+PSO, and Circle+PSO, for image segmentation through multilevel thresholding on the Berkeley Segmentation Dataset (BSDS500). Chaotic sequences generated by the Tent, Sinusoidal, and Circle maps replaced the pseudo-random number generators conventionally used for population initialisation, inertia-weight adjustment, acceleration-coefficient modulation, and particle perturbation. Five hundred BSDS500 images were preprocessed, and the three variants were implemented in Python and evaluated over 30 independent runs per algorithm using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Feature Similarity Index (FSIM), segmentation accuracy, and F1-score. Tent+PSO achieved a PSNR of 26.34 ± 1.02 dB, an SSIM of 0.9056 ± 0.0212, and an F1-score of 0.9362 ± 0.0080; Sinusoidal+PSO achieved a PSNR of 26.89 ± 0.95 dB, an SSIM of 0.9123 ± 0.0198, and an F1-score of 0.9411 ± 0.0074; and Circle+PSO achieved a PSNR of 26.12 ± 1.05 dB, an SSIM of 0.9012 ± 0.0224, and an F1-score of 0.9311 ± 0.0083. Sinusoidal+PSO recorded the highest mean value on all the five evaluation metrics. The study establishes that, for multilevel-thresholding image segmentation on natural images of the kind represented in BSDS500, the Sinusoidal map is the most effective of the three chaotic maps evaluated, followed by the computationally inexpensive Tent map, with the two-parameter Circle map performing least effectively despite its additional tunability.

Keywords: Image Segmentation, Multilevel Thresholding, Chaos-Enhanced Particle Swarm Optimisation, BSDS500.