Authors

Abstract

To accurately detect densely distributed kiwifruit flowers in complex orchard environments, this study proposes an improved detection model, YOLOv11-TYW, based on the YOLOv11n architecture. First, the RepViTBlock is integrated to enhance the model’s feature representation capabilities. Second, the ADown module is introduced to improve the downsampling structure, thereby increasing detection accuracy for small flowers and branches while enhancing inference efficiency. Third, a triple attention module is embedded in the head network to improve detection performance under conditions of occlusion caused by branches and overlapping flowers.Experimental results show that the YOLOv11-TYW model achieves a precision of 88.4%, a recall of 89.1%, and a mean average precision (mAP) of 91.2%, representing improvements of 4.3, 4.4, and 6.7 percentage points, respectively, over the baseline YOLOv11n model. When tested on kiwifruit flower images captured in various orchard environments, YOLOv11-TYW produces more accurate bounding boxes, with fewer false positives and missed detections. These findings demonstrate that YOLOv11-TYW exhibits excellent detection performance in real-world orchard settings and offers technical support for automated kiwifruit flower pollination.

Abstract in Chinese

为实现对果园环境中复杂分布猕猴桃花朵的准确检测，本研究提出一种基于改进YOLOv11n的猕猴桃花朵检测模型YOLOv11-TYW。首先，在YOLOv11n的基础上引入RepViTBlock增强模型的特征表达能力。其次，引入ADown模块改进模型的下采样结构，提升模型对小花朵、枝叶的检测精度与模型推理效率；最后，引入三重注意力模块改进Head网络结构，提升模型对于枝叶遮挡、相互遮挡猕猴桃花朵的检测精度。结果说明，YOLOv11-TYW模型的精确率、召回率和平均精度均值分别为88.4%、89.1%与91.2%，相比 YOLOv11n模型其精确率、召回率和平均精度均值分别提高了 4.3、4.4 和 6.7 个百分点。使用不同环境的猕猴桃花朵照片对改进模型进行检测时，改进的YOLOv11-TYW相较于YOLOv11n模型的预测边界框更接近花朵目标，并减少了误检与漏检的情况。结果表明，YOLOv11-TYW模型在真实猕猴桃果园环境中表现出优良的检测性能，能够实现对密集分布猕猴桃花朵的准确检测，可为猕猴桃花朵的自动授粉提供技术支持。