LIGHTWEIGHT DEEP NEURAL NETWORKS FOR HIGH-ACCURACY RICE DISEASE DETECTION IN RESOURCE-CONSTRAINED ENVIRONMENTS

Abstract

Early and precise detection of rice diseases is vital in preventing losses and promoting food security. In this research, we contrast the performance of three current CNN architectures—DenseNet121, YOLOv8s, and ConvNeXtBase—on a rice disease dataset that contains six different classes. We compare the performance of all models based on vital performance indicators such as classification accuracy, precision, recall, and validation loss. Our experiments demonstrate that DenseNet121 and ConvNeXtBase have competitive accuracy levels of up to 82% and 96% on validation data, but YOLOv8s surpasses them by a large margin by recording a top-1 accuracy of 99% and top-5 accuracy of 100% at with very minimal computational cost. This renders YOLOv8s an ideal candidate for real-time, resource-limited applications like mobile or edge deployment. Contrasted with previous works that usually drew on heavier or less precise models for the same tasks, our findings demonstrate the efficacy of lightweight object detection models such as YOLOv8s for specialized classification contexts. This study not only assists in deepening deep learning methodologies for agricultural diagnosis but also the development of efficient and scalable solutions for deployment in the field.