基于改进CenterNet的小麦条锈病菌夏孢子自动检测方法  

Automatic Detection Method for Urediniospores of Wheat Stripe Rust Based on Improved CenterNet Model

作　　者：雷雨 周晋兵 何东健[3,4] 陈鹏 曾伟辉[1,2] 梁栋 LEI Yu;ZHOU Jinbing;HE Dongjian;CHEN Peng;ZENG Weihui;LIANG Dong(National Engineering Research Center for Agro-Ecological Big Data Analysis and Application,Anhui University,Hefei 230601,China;School of Internet,Anhui University,Hefei 230601,China;College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Key Laboratory of Agricultural Internet of Things,Ministry of Agriculture and Rural Affairs,Yangling,Shaanxi 712100,China)

机构地区：[1]安徽大学农业生态大数据分析与应用技术国家地方联合工程研究中心,合肥230601 [2]安徽大学互联网学院,合肥230601 [3]西北农林科技大学机械与电子工程学院,陕西杨凌712100 [4]农业农村部农业物联网重点实验室,陕西杨凌712100

出　　处：《农业机械学报》2021年第12期233-241,共9页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家自然科学基金项目(62072002);安徽省科技重大专项(202003a06020016);安徽省自然科学基金项目(2108085MC95);农业生态大数据分析与应用技术国家地方联合工程研究中心开放项目(AE2018009)。

摘　　要：针对孢子捕捉设备采集的显微图像中真菌夏孢子自动检测存在严重的误检和漏检问题,提出一种基于改进CenterNet的小麦条锈病菌夏孢子自动检测方法。首先,针对夏孢子显微图像孢子目标微小、种类少等特点,通过减半Basic Block层数,优化CenterNet网络中的特征提取网络,提高了检测和训练速度,降低了误检率;其次,根据孢子形态为近椭圆或圆形的特点,将原始用于CenterNet训练的目标长宽,改进为目标的椭圆框长短轴长度和角度,提高了孢子分割重合率;最后,提出使用椭圆的长短轴映射矩形来计算椭圆框热图的高斯核半径,以减少孢子的漏检率。实验结果表明,改进的CenterNet夏孢子检测方法对小麦条锈病菌夏孢子检测的识别精确率达到了98.77%,重叠度为83.63%,检测速度为41 f/s,达到了实时检测的应用需求,比原始的CenterNet模型重叠度提高了7.53个百分点,检测速度快11 f/s,模型占用内存降低了68.5%。本文方法能够精准检测并分割出显微图像中的夏孢子,可为农田空气中小麦条锈病菌夏孢子的自动检测及条锈病的早期预防控制提供技术支持。An automatic detection method for urediniospores of wheat stripe rust based on improved CenterNet model was proposed to solve the serious problems of false detection and missing detection in microscopic images collected by spore capture equipment.Firstly,the feature extraction network in CenterNet was optimized by halving the number of Basic Block layers to improve detection and training speed in view of the characteristics of small spore targets and few species in fungal spore microscopic images.Secondly,according to the characteristic that the spore shape was nearly elliptic or round,the width and height of the target originally used for training was improved into the long and short axis and angle of the target ellipse frame for the training part of CenterNet,which improved the spore segmentation coincidence rate.Finally,the long and short axis mapping rectangle was used to calculate the Gaussian core radius of the elliptical frame heat map to reduce the missed detection rate of spores.The experimental results showed that the identification accuracy of the improved CenterNet detection method for urediniospores of wheat stripe rust was 98.77%,the overlap degree was 83.63%,and the detection speed was 41 f/s,which met the application requirements of real-time detection.Compared with the original CenterNet model,the overlap degree was increased by 7.53 percentage points,the detection speed was 11 f/s faster,and the model size was reduced by 68.5%.In conclusion,the experimental results indicated that the proposed method can accurately detect and segment fungal spores in the microscopic image,providing technical support for the automatic detection of airborne urediniospores of wheat stripe rust in wheat fields and the early control of wheat stripe rust.

关 键 词：小麦条锈病 真菌夏孢子 自动检测 改进CenterNet 深度学习 椭圆框 

分 类 号：S435.121.42[农业科学—农业昆虫与害虫防治] TP391.4[农业科学—植物保护]