复杂背景下多种树形果树树冠的检测与识别方法研究  

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

作　　者：肖珂[1,2] 王文静 高冠东 么炜[1] XIAO Ke;WANG Wenjing;GAO Guandong;YAO Wei(College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China;Hebei Key Laboratory of Agricultural Big Data,Baoding 071001,China;Department of Information Management,The National Police University for Criminal Justice,Baoding 071000,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071001 [2]河北省农业大数据重点实验室,河北保定071001 [3]中央司法警官学院信息管理系,河北保定071000

出　　处：《河北农业大学学报》2022年第4期100-108,共9页Journal of Hebei Agricultural University

基　　金：国家自然科学基金项目(31801782);河北省自然科学基金项目(C2020204055);河北省省级科技计划资助项目(20327401D)

摘　　要：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。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.

关 键 词：果树树冠 改进Mask R-CNN 多种树形 复杂背景 图像检测与识别 

分 类 号：S126[农业科学—农业基础科学]