复杂背景下多种树形果树树冠的检测与识别方法研究

针对多种树形果园环境下，由于树冠背景复杂导致的树冠分割、检测及树形识别困难的问题，本研究提出了1种改进Mask R-CNN的B-Mask R-CNN检测模型，实现自然复杂环境下的果树树冠检测与树形识别。该模型引入了IoU(Intersection over Union)平衡采样，提高了模型的训练效果；其次，在边界框损失中引入平衡L1损失，使得多分类损失与边界框损失更快地收敛；另外，在区域推荐网络中调整锚框比例适应数据集中的目标，提升了模型准确率。该研究搜集矮化密植形、小冠疏层形、自然开心形、自然圆头形以及Y形5种常见修剪树形制作数据集，应用5个检测模型进行对比试验。试验结果表明，B-Mask R-CNN检测模型平均检测精度达到98.7%，与Mask R-CNN、Faster R-CNN、U-Net以及K-means模型相比检测精度更高，对复杂背景下的树形识别具有更好的鲁棒性，能够为后续精准喷施中喷施模式和控制参数的分析及应用奠定基础。

Study on crown detection and recognition method of multiple fruit tree shapes under complex background#br#

The complex tree shapes and backgrounds usually influence the accuracy of fruit tree shape recognition,which is an important step on precise spraying, automatic trim and picking. In order to solve these problems, deep learning was introduced to improve the recognition accuracy in recent years. In this paper, a crown recognition method based on B-Mask R-CNN network was proposed to detect and recognize multiple shapes of fruit trees in the complex environment. Firstly, IoU-balanced sampling was introduced, which improved the training effects of the model. Then, Balanced L1 loss was introduced to the bounding-box loss to make loss converge faster. In addition, the proportion of Anchor from Region Proposal Network was adjusted to fit the targets in data set, which improved the accuracy of the model. Finally, a crown data set of fruit trees containing five common pruning tree shapes was built and five detection models were used for comparative experiments. The results showed that the mAP of B-Mask R-CNN reached 98.7%. It achieved better detection and segmentation effect compared with Mask R-CNN, Faster R-CNN,U-Net and K-means. In summary, this study provided an effective method for the detection and recognition of various fruit tree shapes in the complex background and established a foundation for the analysis and application of spraying mode and control parameters in subsequent precision spraying.