Image Captioning with PSO-Optimized ResNeXt and Custom Transformer Architecture

Image captioning is a crucial task in computer vision and natural language processing, requiring effective feature extraction and contextual understanding. Traditional convolutional and Transformer-based architectures face challenges in selecting optimal model configurations, often leading to suboptimal performance. To address these issues, we propose a novel approach that integrates Particle Swarm Optimization (PSO) with a ResNeXt- based CNN encoder and a custom Transformer-based decoder. PSO is employed to optimize the number of ResNeXt blocks and cardinality, ensuring an efficient and automated architectural selection process that enhances feature extraction. Additionally, we design a proprietary Transformer architecture incorporating multiple multi-head attention mechanisms, enabling the model to attend to diverse relevant image regions simultaneously. This significantly improves contextual understanding, leading to more informative and accurate caption generation. Experimental results demonstrate that our approach achieves state-of-the-art performance, with BLEU-1, BLEU-2, BLEU-3, and BLEU-4 scores of 79.17, 55.64, 42.79, and 34.23, respectively, surpassing existing methods. These findings highlight the effectiveness of PSO-optimized ResNeXt and custom Transformer architectures in advancing image captioning performance.