Advanced Swarm Intelligence For Image Processing

Abstract

Swarm intelligence algorithms are inspired by natural behaviors and phenomena, such as those observed in insects, birds, and other biological populations. These algorithms mimic natural processes to solve complex problems through intelligent computing methods. They are widely recognized for their effectiveness in image processing. This thesis explores their potential and proposes enhancements to improve their performance in this field. This thesis begins with an in-depth review of leading swarm intelligence optimization algorithms such as Ant Colony Optimization (ACO), Gravitational Search Algorithm (GSA), and Particle Swarm Optimization (PSO). Focusing on their use in image processing, we assess their advantages and limitations through a detailed comparative analysis. A primary focus is on improving ACO for image edge detection, which traditionally struggles with local optima, noise handling, and parameter sensitivity. A Modified Ant Colony Optimization (MACO) algorithm is introduced to address these issues. MACO enhances ant distribution, uses a new probability transfer function, and introduces a dynamic pheromone threshold for adaptability. Experimental results show that MACO produces clearer, more continuous edges comparing to the other algorithms. The thesis also presents a Hybrid Gravitational Search Algorithm (HGSA), which combines GSA’s data utilization with PSO’s fast convergence to optimize deep neural network hyperparameters, enhancing face recognition accuracy and robustness. PSO, key to HGSA's success, is highly efficient but faces challenges like premature convergence and loss of diversity. To address these limitations, three PSO variants were developed and applied to object tracking. Experimental results show significant improvements in both accuracy and applicability. This underscores the potential of these algorithms to drive forward technological advancements in image processing.