Enhancing Image Classification: A Metaheuristic-Driven Approach

Image classification plays an important role in many domains, ranging from healthcare to autonomous systems. Gaining high accuracy and optimal performance in image classification tasks profoundly relies on fine-tuning the hyperparameters and architecture configurations of Convolutional Neural Networks (CNNs). In the current study, we propose a metaheuristic-driven technique to optimize and adjust the hyperparameters and CNN architectural configurations binding the capabilities of Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). By using the search capabilities of GA and PSO, our method automates the process of detecting optimal settings, removing the need for manual trial and error. The GA and PSO algorithms permit concurrent exploration of the hyperparameter space and architectural choices of the CNN model. Through the experiments on benchmark datasets, we establish the efficacy of our approach in refining image classification performance. The proposed study of improving hyperparameter for CNN architecture resulted in an accuracy of 93.22 and 95.45% for CIFAR-10 dataset and 98.36 and 97.91% for MNIST dataset for GA and PSO, respectively. This work highlights the capacity of metaheuristic algorithms as powerful tools for optimizing CNNs in image classification tasks, providing a capable avenue for future research and applications.