IP-YOLO: Progressive Fusion and Deep Vision Transformer for Accurate Insect Pest Detection

Insect pests pose a significant threat to agricultural productivity, necessitating accurate and timely identification to minimize economic losses. This paper presents IP-YOLO, an enhanced insect pest detection algorithm that leverages the YOLOv8 framework. IP-YOLO introduces several key enhancements to address the challenges of traditional algorithms, particularly in recognition accuracy and response speed. First, EfficientViT is utilized as the backbone network to efficiently extract deep image features and capture global information. Second, an Asymptotic Feature Pyramid Network (AFPN) is introduced to reduce feature loss during multi-scale fusion, thereby improving detection accuracy. Furthermore, the C2fSE module is proposed to enhance the algorithm's image-capturing ability while minimizing noise and redundancy. Extensive experiments on the IP50 dataset demonstrate that IP-YOLO outperforms several popular baselines, improving P by 5%, R by 6%, and mAP50 by 4.1% compared to the original YOLOv8 algorithm. These results indicate that IP-YOLO is a promising candidate for real-time monitoring and control of insect pests, offering a practical solution for high-quality pest detection with reduced computational cost.