Unlocking the power of CNN models: enhancing training procedure with evolutionary based search for class-specific data augmentation

Data augmentation (DA) is a critical technique for improving the generalization capa-bilities of Convolutional Neural Networks (CNNs) in image classification tasks. This re-search introduces an automated data augmentation framework that leverages a genetic algorithm (GA) to optimize class-specific augmentation strategies, aiming to enhance CNN performance across diverse datasets and architectures. Our framework integrates a GA to search for optimal DA strategies and a CNN as a fitness function to evalu-ate the effectiveness of each strategy. By employing this fitness-driven, evolutionary approach, the framework iteratively refines augmentation strategies over generations, selecting individuals that yield improved model performance. The framework was rigor-ously evaluated on a diverse set of datasets spanning medical, agricultural, and general domains. Results demonstrated significant improvements in validation loss, with reduc-tions of up to 58.56% compared to a non-augmented baseline and 54.28% compared to a baseline augmented with traditional methods. Additionally, our approach was benchmarked against a similar state-of-the-art framework from the literature. In this comparison, our method achieved a relative improvement of 210%, underscoring its su-perior efficiency in optimizing DA strategies. These findings highlight the framework’s robustness and potential to serve as a universal solution for enhancing CNN classifi-cation performance. The framework provides a practical, adaptable tool that can be seamlessly integrated into existing workflows to improve model performance, particu-larly in domains with limited data and contributes to the growing field of automated data augmentation.