Evaluation of Deep Learning Models for Analyzing the Severity of Rice Grain Damage Induced by Varying Threshing Conditions

Mechanical threshing significantly contributes to postharvest losses in rice production by inducing physical damage that compromises grain quality and market value. Traditional visual assessment methods for grain damage are subjective, labor-intensive, and inefficient at scale. This study presents an Artificial Intelligence (AI)-based approach for classifying and analyzing rice grain damage caused by two threshing element types, rasp bars and elbow rod teeth, under controlled experimental conditions. A total of 8133 high-resolution RGB images were acquired using an industrial imaging system, with grains manually categorized into four classes: completely broken, partially broken, whole with minor damage, and healthy. The images were preprocessed and augmented using rotation, flipping, and scaling to enhance model generalization. Five state-of-the-art deep learning models, such as Vision Transformer (ViT-B-16), ResNet-50, EfficientNet-B0, MobileNetV2, and MobileNetV3-Small, were implemented using transfer learning and fine-tuned on the rice damage data set. Models were trained on stratified data sets and evaluated using metrics such as accuracy, precision, recall, F1-score, and validation loss. The ViT-B-16 and ResNet-50 models achieved the highest classification accuracies of 98.31% and 98.36%, respectively, with low validation losses (0.05 and 0.07), indicating strong generalization. Grad-CAM and attention map visualizations were used to interpret the decision-making processes of ResNet-50 and ViT-B-16, respectively. Comparative analysis showed that rasp bar threshing elements induced significantly more severe damage than elbow rod teeth. The proposed methodology demonstrates that deep learning models, particularly ViT-B-16 and ResNet-50, offer a robust, objective, and scalable solution for evaluating rice grain damage. This AI-powered approach provides a valuable tool for optimizing threshing processes, enhancing postharvest quality control, and reducing economic losses in rice production systems.