WSD-YOLO: A lightweight YOLO-based model with enhanced feature representation for maize pest detection

Abstract

As a major staple crop with global significance, maize is highly vulnerable to pest infestations throughout its growth cycle, which can substantially limit plant development and reduce yield. In practice, accurate detection remains challenging due to pronounced morphological variations across pest life stages, as well as interference from complex field backgrounds, often leading to missed detections. To address these challenges, we propose WSD-YOLO (YOLOv11n with WindmillConv, Single-Scale High-Level Transformer, and Dual-Attention Weighted Aggregation), a high-accuracy maize pest detection model built upon an improved YOLOv11n framework. The WindmillConv (WMConv) module enhances the model’s sensitivity to multi-directional pest textures, thereby improving low-level feature representation while maintaining a lightweight convolutional design. At a deeper level, the Single-Scale High-Level Transformer (SHLT) introduces global self-attention, enabling effective suppression of background noise with minimal computational overhead. In addition, the Dual-Attention Weighted Aggregation (DAWA) module adaptively fuses same-scale features, improving the detection of pests with diverse morphological characteristics. Experimental results from three repeated training runs (random seeds: 42, 2026, 3407) on the IPMaize dataset demonstrate that the proposed method achieves an averaged mAP@0.5 of 78.77% (79.7% with seed = 42) and averaged mAP@0.5:0.95 of 52.63%. Cross-dataset evaluations on the Tomato Pest&Diseases and IP102 datasets further confirm the model’s strong generalization capability.