Pest-YOLOv8: A Novel Lightweight Algorithm for Pest Detection

Intelligent pest detection is a crucial application of object detection in agriculture, significantly improving the efficiency of pest monitoring and ensuring crop quality and yield. However, current pest detection algorithms suffer from low accuracy and large parameter scales. To address these issues, this paper proposes a lightweight pest detection algorithm named Pest-YOLOv8. Firstly, a novel lightweight feature extraction module is introduced based on FasterNet. Secondly, a new lightweight detection head is proposed on the foundation of the YOLOv8 detection head, significantly reducing model parameters and computational complexity while maintaining detection accuracy. Subsequently, the neck portion of the original YOLOv8 is reconstructed using the SimAM attention mechanism, noticeably enhancing the model's feature extraction capabilities in complex natural environments. Finally, a pruning strategy based on layer-adaptive magnitude is designed according to the YOLOv8 network structure, drastically reducing the model size while ensuring performance, making it more suitable for deployment on edge devices. Experimental results on the IP102 dataset subset demonstrate that the Pest-YOLOv8 model achieves a mAP0.5 of 96.6% and a recall rate of 94.5%, outperforming classic models such as YOLOv8n, SSD, and Faster-RCNN. Compared to YOLOv8n, Pest-YOLOv8 shows a 1.4% improvement in mAP0.5 and a 2.4% increase in recall, with a reduction of 57.1% in parameters and 65.5% in computational complexity. This indicates that the improvements in this paper effectively reduce instances of missed and false detections, providing a more efficient solution for deploying models on edge devices like cameras, thereby enhancing pest monitoring, early warning, and prevention services.