复杂棉田目标检测失效机理与模型轻量改进

Failure Mechanisms and Lightweight Model Improvement for Object Detection in Complex Cotton Fields

关键词:  棉花  YOLO  轻量化模型  目标检测  棉田管理

Key words:  cotton  YOLO  lightweight model  object detection  cotton-field management

摘要:

【目的】针对复杂棉田中棉花花朵、未开放棉铃和开放棉铃在光照变化、枝叶遮挡、目标粘连及尺度变化条件下易漏检、误检的问题，构建兼顾检测精度与推理效率的轻量化模型，为棉花生育监测和打落叶剂施用时机辅助决策提供视觉基础。【方法】构建包含3类棉花生育目标的数据集，图像覆盖不同生育时期、拍摄角度、光照条件和田间背景。对多种YOLO轻量模型进行比较，并以YOLOv8n为基础提出RepAgri-YOLO，构建重参数化卷积模块RepC2f、内容感知门控上采样模块CAGU和动态可学习加权拼接模块DLWC，同时设计召回增强联合损失函数REJLoss，以增强方向性特征提取、跨尺度特征融合和困难正样本学习能力。【结果】（1）原始模型中YOLO26n综合性能较优，mAP50为0.75。RepAgri-YOLO在总测试集上的召回率和mAP50分别为0.64和0.76，较YOLOv8n分别增加0.09和0.04，计算量GFLOPs由8.1 G降至6.9 G。其中，未开放棉铃的召回率由0.46提高至0.61，mAP50由0.67提高至0.75。（2）消融实验表明，RepC2f使模型计算量GFLOPs下降了18.2%的同时保持了模型精度和召回率，随着GAGU模块加入，模型的整体召回率和mAP50分别由0.55和0.72提升至0.64和0.76，而DLWC和REJLoss则表现出协同作用，最终RepAgri-YOLO相对YOLOv8n在降低14.18%计算量的同时使得召回率和mAP50分别提升了16.36%和5.56%。【结论】复杂棉田检测的主要困难为未开放棉铃与背景对比度低、开放棉铃密集粘连及跨尺度融合不充分。RepAgri-YOLO在较低计算量下提高了困难目标检出能力，尤其缓解了未开放棉铃漏检问题，可为棉田生育状态监测和智能化管理提供技术支撑。

Abstract:

[Purpose] Cotton flowers, unopened cotton bolls, and opened cotton bolls in complex cotton-field environments are prone to missed and false detections under illumination variations, branch and leaf occlusion, target adhesion, and scale changes. This study aims to develop a lightweight detection model that balances detection accuracy and inference efficiency, thereby providing a visual basis for cotton growth monitoring and auxiliary decision-making on the timing of defoliant application. [Method] A dataset containing three categories of cotton growth-related targets was constructed, with images covering different growth stages, shooting angles, illumination conditions, and field backgrounds. Several lightweight YOLO models were compared, and RepAgri-YOLO was developed based on YOLOv8n. A reparameterized convolution module (RepC2f), a content-aware gated upsampling module (CAGU), and a dynamic learnable weighted concatenation module (DLWC) were constructed. In addition, a recall-enhanced joint loss function (REJLoss) was designed to enhance directional feature extraction, cross-scale feature fusion, and the learning of difficult positive samples. [Result] (1) Among the original models, YOLO26n achieved relatively good overall performance, with an mAP50 of 0.75. On the overall test set, RepAgri-YOLO achieved a recall of 0.64 and an mAP50 of 0.76, representing absolute increases of 0.09 and 0.04, respectively, compared with YOLOv8n, while the computational complexity decreased from 8.1 to 6.9 GFLOPs. In particular, the recall for unopened cotton bolls increased from 0.46 to 0.61, and the mAP50 increased from 0.67 to 0.75. (2) Ablation experiments showed that RepC2f reduced the computational complexity by 18.2% while maintaining model accuracy and recall. After the CAGU module was introduced, the overall recall and mAP50 increased from 0.55 and 0.72 to 0.64 and 0.76, respectively. DLWC and REJLoss exhibited a synergistic effect. Consequently, compared with YOLOv8n, the final RepAgri-YOLO reduced computational complexity by 14.18% while improving recall and mAP50 by 16.36% and 5.56%, respectively. [Conclusion] The main challenges in complex cotton-field detection include the low contrast between unopened cotton bolls and the background, dense adhesion of opened cotton bolls, and insufficient cross-scale feature fusion. RepAgri-YOLO improves the detection capability for difficult targets with relatively low computational complexity, particularly alleviating missed detections of unopened cotton bolls. The proposed model provides technical support for cotton growth-status monitoring and intelligent cotton-field management.