Efficient rice pest intelligent detection framework for field environments

Abstract

Rice pests pose a serious threat to both the yield and quality of crops. Traditional detection methods are inefficient, subjective, and fail to meet the demands of modern agriculture for intelligent, real-time, and high-precision detection. To address the challenges of existing lightweight object detection models, such as inadequate multi-scale recognition and small object detection ability in complex pest environments, this paper proposes a rice pest detection model based on an improved YOLOv11n. The model introduces the C3k2-ConvFormer-CGLU backbone structure, which integrates local convolution and a lightweight attention mechanism to enhance fine-grained feature expression. It also designs the ContextGuideFPN module, which optimizes multi-scale feature fusion through context guidance, and constructs the C2DAttention module, enabling dynamic attention to key areas and redundancy suppression. Trained and tested on a self-built dataset containing six typical rice pest species, experimental results demonstrate that the proposed model outperforms existing mainstream lightweight models in terms of accuracy, recall rate, and mAP. With an mAP@0.5 of 95.5%, the model achieves an optimal balance between detection performance and deployment efficiency while maintaining a low parameter count (2.49M) and high inference speed (416.6 FPS). The model demonstrates strong adaptability and practical application potential in complex pest detection scenarios.