YOLO-DCPG: A Lightweight Architecture with Dual-Channel Pooling Gated Attention for Intensive Small-Target Agricultural Pest Detection

Jingwei Liu^1,2,3En Yu^1,2,3Yongke Li^1,2,3*Yunjie Zhao^1,2,3*Bowen Mao^1,2,3

• ¹Xinjiang Agricultural Informatization Engineering Technology Research Center, Urumqi, China
• ²College of Computer and Information Engineering, Xinjiang Agricultural University, Urumqi, China
• ³Research Center for Intelligent Agriculture, Ministry of Education Engineering, Urumqi, China

Accurate and rapid identification of agricultural pests using deep learning models is crucial for intelligent pest monitoring in agriculture. However, existing pest detection models often suffer from high parameter counts and computational complexity, which significantly increases hardware requirements and limits their deployment on edge devices. To address these challenges, a lightweight agricultural pest detection model named YOLO-DCPG is proposed. It is built upon YOLOv8n. First, an attention module called Dual Channel Pooling Gated Attention (DCPGAttention) is designed. This module applies mean and standard deviation pooling to the input features to enhance global information capture. Second, a lightweight backbone network, StarNet, is used for feature extraction. A feature fusion neck, called Small-Neck, is also introduced. It is based on an improved bidirectional feature pyramid network (a-BIFPN) and integrates the efficient GSConv module. This design reduces model parameters and computational cost while maintaining detection accuracy. Finally, a scale factor based on the Inner-IoU is embedded into the WIoU loss function, enabling more precise control over auxiliary bounding boxes. This allows the model to focus more effectively on small pest regions and accelerates the regression of small target sample. The experimental results show that the proposed method achieves an precision of 80.1%, mAP@50 of 74%, and mAP@50~95 of 47.5% on the Pest24 dataset. In comparison with the baseline model YOLOv8n, it improves by 4.5%, 0.8%, and 0.9%, respectively. Meanwhile, the number of parameters, giga floating-point operations per second (GFLOPs), and model size are reduced by 51.2%, 30.1%, and 46.7%, respectively. Finally, YOLO-DCPG is successfully deployed on the Raspberry Pi 4B hardware, achieving stable and real-time pest detection.