| 罗兴发,潘良,李磊,董继伟,湛小梅,李亚丽.基于改进YOLOv11s的含遮挡柑橘检测研究[J].中国南方果树,2026,55(4):255-262 |
| 基于改进YOLOv11s的含遮挡柑橘检测研究 |
| Research on Occluded Citrus Fruit Detection Based on Improved YOLOv11 Algorithm |
| 投稿时间：2025-07-11 修订日期：2025-08-11 |
| DOI： |
| 中文关键词: 重叠柑橘 柑橘采摘机器人 计算机视觉 目标检测 深度学习 YOLOv11 |
| 英文关键词:Overlapping citrus Citrus harvesting robot Computer vision Object detection Deep learning YOLOv11 |
| 基金项目:重庆市农业科学院科技创新类项目基于机器视觉的柑橘采摘障碍识别与重建方法研究（KYLX20240500074）、重庆市科研机构绩效激励引导专项面向非结构环境的柑橘低损抓取及采摘路径优化研究（CSTB2024JXJL-YFX0004） |
|
| 摘要点击次数: 9 |
| 全文下载次数: 0 |
| 中文摘要: |
| 【目的】为解决在自然环境下对含遮挡柑橘果实检测精度低、实时性差的难题，提出一种基于YOLOv11改进模型的柑橘识别方法。【方法】首先，引入RFEM(递归特征增强模块)，通过多分支空洞卷积融合多尺度上下文信息，提升对遮挡目标的特征提取能力；其次，在检测头嵌入EVCBlock(高效视觉卷积块)，利用动态加权和轻量MLP强化边缘特征建模，并重构特征金字塔的跨层级特征融合路径；最后，解耦CIoU损失函数的宽高比惩罚项，优化小目标定位精度，并针对遮挡和质量缺陷做参数调优。【结果】实验采用29869张自然场景柑橘图像作为数据集，结果显示：改进模型mAP50达93.3%，F1分数达到0.88，检测速度达52 FPS，相较于当前热门的通用目标检测模型YOLOv8s-world和RT-DETR，展现出更高的综合性能优势。【结论】本文提出的基于YOLOv11的改进方法有效平衡了遮挡场景下的检测精度与实时性，为果园采摘机器人的实用化部署提供了可靠的技术支撑。 |
| 英文摘要: |
| [Purpose] To address the challenges of low detection accuracy and poor real-time performance for occluded citrus fruits in natural environments, this paper proposes an improved model based on YOLOv11 for citrus recognition. [Methods] Firstly, the RFEM (Recursive Feature Enhancement Module) is introduced to fuse multi-scale contextual information via a multi-branch architecture with dilated convolutions, thereby enhancing feature extraction capabilities for occluded targets. Subsequently, the EVCBlock (Efficient Visual Convolution Block) is embedded into the detection head to enhance edge feature modeling via dynamic weighting and a MLP (lightweight Multilayer Perceptron), and reconstructs the cross-level feature fusion pathways in the feature pyramid. Finally, the approach decouples the aspect ratio penalty term in the CIoU loss function to enhance the localization accuracy of small targets, with parameter tuning specifically designed to address occlusion and quality defects. [Results] Experiments conducted on a dataset of 29,869 natural-scene citrus images demonstrate that the improved model achieves 93.3% mAP50, an F1-score of 0.88, and a detection speed of 52 FPS. Compared to state-of-the-art general object detection models like YOLOv8s-world and RT-DETR, the proposed approach demonstrates superior overall performance advantages. [Conclusion] This enhanced YOLOv11-based methodology effectively balances detection accuracy with real-time performance in occlusion scenarios, providing robust technical support for practical deployment in orchard harvesting robots. |
| 查看/发表评论 下载PDF阅读器 |
| 关闭 |
