Bi-DS-YOLOv8:基于YOLOv8n的轻量化田间小麦麦穗检测模型OA

[Objective]Field wheat spike detection faces challenges such as target occlusion,low resolution,and target blurring caused by complex field environments,which lead to insufficient detection accuracy of existing models.To address these issues,the present study propo-ses a targeted improvement scheme based on the YOLOv8n model.[Method]First,a deformable attention mechanism was introduced into the C2f module to enhance the model's ability to extract features of deformed targets.Second,SPDConv was used to replace all strided convolu-tional layers in the backbone network to avoid the loss of small-target information.Finally,the PANet feature fusion structure was optimized to BiFPN to achieve efficient weighted fusion of multi-scale features.To verify the effectiveness of the comprehensive improvement,comparative experiments were conducted:three single-improvement models(introducing only deformable attention,replacing only SPDConv and optimizing only BiFPN)were compared with the proposed comprehensive improved model in terms of detection performance.[Result]The comprehen-sive improved model achieved significant performance gains in field wheat spike detection.In terms of accuracy,the mAP@0.5 of the im-proved model reached 93.9％,a 1.1 percentage point increase compared with the original YOLOv8n;the mAP@0.95 reached 58.7％,a 1.3 percentage point improvement.Both precision and recall were improved simultaneously,while the model maintained a high frames per sec-ond(FPS),satisfying the real-time detection requirement.In the single-improvement experiments,the mAP@0.5 values of the models with only deformable attention,only SPDConv replacement,and only BiFPN optimization were 92.4％,92.1％and 92.7％,respectively-all low-er than the 93.9％of the comprehensive improved model,confirming the synergistic enhancement effect of the three improvement methods.[Conclusion]The comprehensive improvement scheme proposed in the study effectively solves the problems of occlusion,low resolution,and target blurring in field wheat spike detection,and its detection accuracy is significantly better than that of the original YOLOv8n and single-improvement models.The achievement provides a reference technical pathway for field crop phenotyping monitoring and the development of vision systems for intelligent agricultural equipment,and holds practical application value for promoting the development of intelligent agricul-tural monitoring terminals.