Abstract
Rice pest detection faces critical challenges including small target recognition difficulties, high morphological similarities, and complex field backgrounds. This study proposes BEAM-YOLO (Bi-branch Edge Attention Multi-scale YOLO) to address these limitations.We constructed the JRICE-PD dataset encompassing 11 economically significant rice pests (4,565 images) and developed four innovative modules: a Multi-scale morphological Edge Network (MEN) enhancing feature discrimination; a Bi-branch Attention Feature Enhancement (BAFE) module utilizing Haar wavelet transform for foreground-background separation; an Enhanced Multi-scale Bidirectional Feature Pyramid Network (EM-BFPN) optimizing information interaction; and a Spatial-Channel Augmented Upsampling (SCAU) improving small target detection.BEAM-YOLO achieves 86.6±0.5% mAP@50 and 72.7±0.9% mAP@50-95, outperforming YOLOv11 by 3.3% and 3.0% respectively, while maintaining relatively low computational overhead and parameter count. This research provides reliable algorithmic support for intelligent agricultural pest monitoring systems, contributing to precision agriculture advancement and application.