基于Mask R-CNN模型的虫害识别方法研究  

Research on Pest Recognition Method Based on Mask R-CNN Model

作　　者：张耀丽 许宁 宋裕民 康建明 张国海[1] 孟庆山 Zhang Yaoli;Xu Ning;Song Yumin;Kang Jianming;Zhang Guohai;Meng Qingshan(College of Agricultural Engineering and Food Science,Shandong University of Technology,Zibo 255000,Shandong,China;Shandong Academy of Agricultural Machinery Sciences,Jinan 250131,Shandong,China;School of Automotive Engineering,Shandong Jiaotong University,Jinan 250357,Shandong,China)

机构地区：[1]山东理工大学农业工程与食品科学学院,山东省淄博市255000 [2]山东省农业机械科学研究院,山东省济南市250131 [3]山东交通学院汽车工程学院,山东省济南市252100

出　　处：《农业装备与车辆工程》2023年第1期44-47,52,共5页Agricultural Equipment & Vehicle Engineering

基　　金：山东省重大科技创新工程项目(2019JZZY020623);山东省重点研发计划项目(2022CXGC020706)。

摘　　要：为提高果园虫害远程监测自动识别的准确率与稳定性,以梨小食心虫作为研究对象,提出了基于Mask R-CNN目标检测模型的虫害识别方法。在Faster R-CNN模型基础上优化添加用以预测分割的Mask回归分支,将进行2次量化操作的Ro I Pooling层替换为使用线性插值算法的Ro I Align层,另外添加并列的FCN层,使模型具有更强大的泛化能力,可在有效检测目标的同时输出高质量的实例分割Mask。采用固定式物联网捕虫装置获取数据样本,并应用数据增广和掩模标注等方法构建数据集,经过深度学习训练,可有效提高果园虫害识别的准确性。实验表明,该方法的目标识别平均准确率达98.87%,能够对梨小食心虫进行精确、稳定、高效的识别,可为精确虫害防治提供参考。In order to improve the accuracy and stability of automatic identification of remote monitoring of orchard pests, a pest recognition method based on Mask R-CNN target detection model is proposed with Grapholita molesta as the research object. The Faster R-CNN model is optimized by adding a Mask regression branch for predictive segmentation, replacing the RoI Pooling layer that performs two quantization operations, with a RoI Align layer, which uses a linear interpolation algorithm. And add a parallel FCN layer, which gives the model a more powerful generalization capability and can output high-quality instance segmentation Masks while effectively detecting targets. The fixed IoT pest trapping device is used to obtain data samples, and the data set is constructed by applying methods such as data augmentation and mask labeling, which can effectively improve the accuracy of orchard pest identification after deep learning training. After experiments, it was shown that the average accuracy of the target recognition of the method reached 98.87%,which can provide accurate, stable and efficient recognition of Grapholita molesta, and can provide reference for precise pest control.

关 键 词：虫害识别 Mask R-CNN模型 样本采集 图像预处理 梨小食心虫 

分 类 号：S126[农业科学—农业基础科学] S435.112