In this paper, a visual detection method by a UAV (unmanned aerial vehicle) was proposed to detect green mangoes on the surface of the tree crown rapidly and meet the need of estimating the number of mango fruits in orchards. The YOLOv2 model was used for quick detection of mango images captured by a UAV. First, mango images were collected by a UAV. These images were marked manually and used to build a training set and a test set. The parameters of the model were determined by experiments. The resolution of the images was 544 × 544 pixels. The batch size was 64, and the initial learning rate was 0.01. The mAP (mean average precision) of the trained model on the training set was 86.4%. Good detection results were achieved on images containing different fruit numbers and different lighting conditions with a precision of 96.1% and a recall rate of 89.0%. Finally, an experiment was conducted to estimate the actual number of green mango fruits. A method of generating an image of the whole mango tree was designed. The fruit numbers estimation model for green mango was obtained by linear fitting between the actual number and the detected number of mangoes. From the comparison of the fruit numbers of 10 mango trees determined by manual calculation and model prediction, an estimation error rate of 1.1% was achieved. The result demonstrated that the algorithm was effective for green mango detection and provided a methodological reference for quick estimation of the number of green mango fruits in orchards.

中文翻译：

---

基于深度学习的果园无人机视觉检测青芒果

本文提出了一种无人机视觉检测方法来快速检测树冠表面的绿色芒果，满足估计果园芒果果实数量的需要。YOLOv2 模型用于快速检测无人机拍摄的芒果图像。首先，芒果图像是由无人机收集的。这些图像被手动标记并用于构建训练集和测试集。模型参数通过实验确定。图像的分辨率为 544 × 544 像素。批量大小为 64，初始学习率为 0.01。训练集上训练模型的 mAP（平均精度）为 86.4%。在包含不同水果数量和不同光照条件的图像上取得了良好的检测结果，准确率为 96.1%，召回率为 89。0%。最后，进行了一个实验来估计绿色芒果果实的实际数量。设计了一种生成整棵芒果树图像的方法。青芒果的果实数量估计模型是通过芒果实际数量与检测数量之间的线性拟合得到的。通过人工计算和模型预测确定的10棵芒果树的果实数量对比，估计误差率为1.1%。结果表明，该算法对青芒果检测是有效的，为快速估计果园中青芒果果实的数量提供了方法学参考。青芒果的果实数量估计模型是通过芒果实际数量与检测数量之间的线性拟合得到的。通过人工计算和模型预测确定的10棵芒果树的果实数量对比，估计误差率为1.1%。结果表明，该算法对青芒果检测是有效的，为快速估计果园中青芒果果实的数量提供了方法学参考。青芒果的果实数量估计模型是通过芒果实际数量与检测数量之间的线性拟合得到的。通过人工计算和模型预测确定的10棵芒果树的果实数量对比，估计误差率为1.1%。结果表明，该算法对青芒果检测是有效的，为快速估计果园中青芒果果实的数量提供了方法学参考。