An adaptive hybrid particle swarm optimization and genetic algorithm approach for hyperparameter tuning in convolutional neural networks applied to fine-grained image classification

Abstract This study presents a novel adaptive hybrid Particle Swarm Optimization–Genetic Algorithm (PSO-GA) for optimizing the hyperparameters of Convolutional Neural Networks (CNNs) in fine-grained image classification tasks. It is a clever idea to equip the global search power of PSO with the local refinement and diversity-conserving signage of the Genetic Algorithm (GA), to prosecute powerful and adaptive confrontations of challenging hyperparameter space. Using this hybrid optimization method, the important CNN parameters, including the number of filters, kernel size, the pooling size, stride length and network depth are optimized dynamically, to ensure maximum classification. Images are pre-processed to a lower resolution of 128 × 128 grayscale in order to have better computational efficiency with minimum loss of feature representation by the pre-process. Every particle in the swarm represents a candidate CNN architecture and its fitness is estimated after testing its with validation. PSO manages international optimization and GA operators which are selection, crossover, and mutation, are stimulated adaptively to discourage early convergence and diversity of the solutions. As experimental evidence shows the adaptive PSO-GA optimized CNN is much more accurate in classifications than typical CNN models including DenseNet-161, ResNet-50, and VGG-19. The given method has an accuracy of 99.56% on Oxford Flowers, 97.45% on Cars, and 95.95% on Dogs, making 25 epochs, which is much less than the 100–150 epochs that are normally required. The hybrid method produced a training accuracy 2.40% better and a validation accuracy 1.99% better than PSO only method on datasets. The newty is in the suggested adaptive formulation, strategy integration, and implementation to the particular problem field. These results showcase the effectiveness, scalability and computational efficiency of the proposed hybrid strategy, which provides a formidable remedy to the optimization of deep learning model structures in computer vision applications in the real world.