Abstract: Based on the YOLOv8 model, the ShuffleNetv2 module and the Conv_MaxPool module were introduced simultaneously to construct the improved YOLOv8 model (YOLOv8-ShuffleNetv2-Conv_MaxPool). By integrating the ShuffleNetv2 module and the Conv_MaxPool module into the YOLOv8 model, the improved YOLOv8 model significantly enhanced the comprehensive performance of rice pests and diseases detection while maintaining its lightweight design, effectively reducing both the false detection rate and the missed detection rate. The improved YOLOv8 model demonstrated excellent performance across multiple datasets, further validating its robustness and generalization ability. Ablation studies demonstrated that, on the custom dataset, compared to the original YOLOv8 model, the improved YOLOv8 model achieved increases of 3.73 percentage points in accuracy, 3.56 percentage points in precision, 3.78 percentage points in recall, and 3.73 percentage points in F1-score, while maintaining a parameter size of only 24.80 MB. On the Coco128 dataset, the improved YOLOv8 model performed the best, with all key metrics averaging approximately 88.00%, significantly outperforming the original YOLOv8 model, the YOLOv8-ShuffleNetv2 model, and the YOLOv8-Conv_MaxPool model. This model effectively enabled rapid and accurate recognition of rice pests and diseases in practical production environments.

Key words: rice pests and diseases, improved YOLOv8 model, lightweight design, recognition model