基于YOLO-GLBW模型的复杂田间环境下玉米田杂草检测OA

To address the issues of mutual occlusion between maize seedlings and weeds,similar morphology,and dense weed distribution in complex field environments,which lead to difficulties in weed target localization,as well as the problems of insufficient long-range dependency modeling,low detection accuracy,and low inference efficiency in existing detection models,this study proposed an improved model,YOLO-GLBW,based on YOLOv8n.First,we embedded the ghost convolution(GhostConv)structure into the C2f modules of both the backbone feature extraction network and all C2f modules of the neck network,which effectively reduced the model's parameter count and computational cost while preserving its full feature representation capabil-ity.We introduced the large separable kernel attention(LSKA)mechanism into the fast spatial pyramid pooling(SPPF)module,which effectively compensated for the model's deficiency in long-range dependency and enhanced its perception capability for small targets.Second,we adopted a bidirectional feature pyramid network(BiFPN)to improve multi-scale target detection and localization performance.Finally,we introduced the wise intersection over union(WIoU)loss function to optimize localization and classification performance in complex scenarios,accelerate the convergence speed,and enhance the generalization ability.In this experiment,we tested the proposed model on a self-constructed dataset of maize seedling and weed images collected from field environments.The results showed that YOLO-GLBW achieved mAP50 of 83.9%,with 2.47 M parameters and a computational cost of 6.3 GFLOPs(1.0×109 floating-point operations).Compared with the original YOLOv8 model,YOLO-GLBW improved precision,recall,mAP50,and mAP50-95 by 1.4 percentage points,2.1 percentage points,1.6 percentage points,and 0.5 percentage points,respectively,while reducing the parameter count and computa-tional cost by 17.9%and 22.2%,respectively.The YOLO-GLBW model significantly reduced computational cost and achieved a lightweight design,while further improving object detection accuracy.In the face of complex scenarios character-ized by dense distribution,morphological similarity,and mutual occlusion between maize seedlings and weeds,the model maintains stable detection accuracy,resolves the trade-off between recognition precision and efficiency,and demonstrates considerable potential for practical field deployment.