Pest-YOLO: A Lightweight Pest Detection Model Based on Multi-level Feature Fusion

Accurate identification of forest and grassland pests is crucial for ecosystem stability and biodiversity. Given the characteristics of pests in forest and grassland environments-such as a wide variety of species, color similarity to the background, small inter-class variability, and large intra-class variability-an improved lightweight target detection model, Pest-YOLO, is proposed. This model addresses the deficiencies of traditional models in terms of recognition accuracy and response speed. Firstly, an online data enhancement method is introduced, applying different image transformation strategies to enable the model to learn more feature representations. Secondly, a multi-level feature fusion method is proposed and combined with the Ghost model to maintain adaptability while compressing the network size. Additionally, the SERes detection head is proposed to improve the network's ability to detect similar targets. Experimental results show that Pest-YOLO achieves the best detection performance on the PEST27 dataset, with 14.16 M fewer parameters and a 3% improvement in the model's mAP0.5 compared to the original YOLOv8 algorithm, demonstrating its effectiveness and potential for forest and grassland pest detection tasks.