基于改进Mask R-CNN的小麦在穗籽粒表型参数测试方法  

A Method for Testing Phenotype Parameters of Wheat Grains on Spike Based on Improved Mask R-CNN

作　　者：王赟赟 李毅念[1] 陈玉仑[1] 丁启朔[1] 何瑞银[1] WANG YunYun;LI YiNian;CHEN YuLun;DING QiShuo;HE RuiYin(College of Engineering,Nanjing Agricultural University,Nanjing 210031)

机构地区：[1]南京农业大学工学院,南京210031

出　　处：《中国农业科学》2024年第12期2322-2335,共14页Scientia Agricultura Sinica

基　　金：国家重点研发计划(2022YFD2300703、2022YFD2300304);中央高校基本科研业务费专项资金资助项目(KYYJ202106)。

摘　　要：【目的】针对小麦籽粒性状参数获取需要脱粒后测量,测量程序繁杂、费时费力的缺点,提出基于深度学习的小麦在穗籽粒表型参数测试方法。【方法】采集镇麦25、宁麦13和农麦88这3个品种小麦穗两侧正视图像,利用小麦穗正视图像构建图像增强数据集,提出深度学习与形态学处理相结合的小麦在穗籽粒表型参数测试方法。首先,建立基于改进Mask R-CNN网络的麦穗颖壳分割模型,模型以ResNet和FNP为特征提取网络并引入坐标注意力(CA)模块、聚合模块和半卷积模块,实现麦穗图像中颖壳的准确定位、分割和籽粒计数。其次,利用分割的麦穗颖壳掩膜图经形态学处理方法提取麦穗颖壳的5个表型参数,并建立麦穗颖壳表型参数与颖壳内籽粒表型参数之间的线性相关关系。最后,利用麦穗颖壳表型参数与籽粒表型参数之间的线性相关关系预测籽粒表型参数。【结果】(1)基于改进Mask R-CNN网络的麦穗颖壳分割模型的平均精确率AP为94.13%,F1值为91.12%,召回率为88.30%,单幅图像平均检测耗时97 ms,可以快速、精准地识别单粒麦穗颖壳。模型的籽粒计数的均方根误差和平均相对误差分别为0.94个和0.65%,可见模型分割籽粒的精度较高。(2)麦穗颖壳与籽粒之间的表型参数粒长、粒厚、面积、周长、长径比线性相关关系式为:y=0.7258x、y=0.5166x、y=0.3748x、y=0.6756x、y=1.4085x,其决定系数(R2)均在0.85以上。(3)利用图像获取的麦穗颖壳参数数据对上述相关性模型进行验证并预测籽粒表型参数,粒长、粒厚、面积、周长、长径比5个参数的均方根误差和平均相对误差分别为0.17 mm、0.08 mm、0.46 mm^(2)、0.33 mm、0.12和0.02%、0.02%、0.02%、0.02%、0.03%,每个参数的预测数据与实际数据之间的决定系数(R2)均在0.85以上,说明本文提出的籽粒表型预测方法可行。【结论】基于深度学习的小麦在穗籽粒表型测试方�【Objective】Wheat grain phenotype parameters were tested after grains only must been threshed by combine,this process was time-consuming,laborious and complicate.Therefore,a method to test morphological parameters of wheat grains on spike based on improved Mask R-CNN was proposed in this research.【Method】Two sides front images of three varieties wheat spikes,including Zhenmai 25,Ningmai 13 and Longmai 88(Early maturity variety),were collected,and then the image enhancement data set was constructed by using Gaussian filter,salt and pepper noise,and vertical flip image enhancement method.A method combined with deep learning and morphological processing for testing phenotype parameters of wheat grains on spike was proposed.Firstly,the improved Mask R-CNN network model for segmenting spike glume was constructed,which was based on feature extraction networks of ResNet and FNP,and the innovative components Coordinate Attention(CA)module,Aggregation module,and Halfconv module were integrated into it.And the glumes on spike image were accurately detected,located,segmented and counted by the improved Mask R-CNN network model.Secondly,five phenotype parameters of the glumes on spike were extracted by using morphologic processing method from the segmented mask image of wheat spike glume,and the linear correlation equations between the phenotype parameters of the glumes and the phenotype parameters of grains were established.Finally,the linear correlation equations between the phenotype parameters of glume and the phenotypic parameters of grain were used to predict the phenotype parameters of grain.【Result】(1)F1 score,average precision(AP)and recall rate of the optimal model for separating spike glume based on the improved Mask R-CNN network model were 91.12%,94.13%and 88.30%,respectively,and the average consuming time for detecting a single image was 97 milliseconds,so the improved Mask R-CNN network model could quickly and accurately identify the glumes on the single wheat spike.The root-mean-square error and av

关 键 词：麦穗 籽粒表型 深度学习 坐标注意力 

分 类 号：S512.1[农业科学—作物学]