<p>Multilevel thresholding is a commonly adopted technique in image segmentation, where an image is partitioned into several meaningful regions. The selection of optimal threshold values is regarded as a difficult problem, and the difficulty is more pronounced for complex and multimodal colour images. A large number of optimization algorithms have been reported for this purpose, although their effectiveness is often constrained by premature convergence, limited population diversity, and the absence of adaptive exploration. An Improved Colony Predation Algorithm (ICPA) is proposed in this paper to mitigate these concerns. Three enhancements are introduced into the behavioural phases of the baseline algorithm. The encircling phase is first supplemented with a quantum tunneling so that stagnation at a local optimum can be avoided. Inter-agent communication is then enriched through an episodic memory pool that retains historically successful positions. A biologically motivated energy fatigue mechanism is finally employed to regulate the search behaviour in accordance with dynamic metabolic constraints. The Minimum Cross Entropy Measure (MCEM) is taken as the objective function for the assessment of segmentation quality. The proposed method is evaluated on images drawn from the Berkeley Segmentation Dataset (BSDS500) and is compared with the Whale Optimization Algorithm (WOA), Aquila Optimization (AO), Parrot Optimization (PO), Particle Swarm Optimization (PSO), and Equilibrium Optimizer (EO). The experimental results indicate that better PSNR, SSIM, and FSIM values are obtained across the tested threshold levels, and a greater degree of statistical stability is observed in comparison with the competing algorithms.</p>