Automated classification of export-grade bananas using CNNs and the cheetah optimization algorithm

A novel ensemble deep learning model proposed for the classification of banana fruit images, primarily to differentiate exportable-quality from rejected samples. A specific dataset, consisting of 1500 balanced images including two banana types generated in ideal circumstances, with the addition of a pre-processing step aimed to exclude the influence of the background. Three custom CNN architectures of increasing complexity (6-layer, 9-layer, and 12-layer) were trained and evaluated using 10-fold cross-validation. In order to improve the classification rate, an ensemble method was designed to tune the classification weights of these CNN models using Cheetah Optimization Algorithm (COA), a new bio-inspired metaheuristic algorithm. The ensemble method outperformed individual CNN models in all metrics and obtained 99.02% accuracy. The results showed that training and transferring large model benefited the performance and the curve of loss function during training and validation phase was able to achieved stable convergence. This research illustrates the promise of COA-based ensemble CNNs for robust, high-accuracy classification in agricultural image analysis, which is applied to food quality inspection and post-harvest automation.