基于改进YOLOv5s的日光温室黄瓜霜霉病孢子囊检测计数方法  

Detection method for cucumber downy mildew sporangia in a solar greenhouse based on improved YOLOv5s

作　　者：李明 丁智欢[1,2] 赵靖暄 陈思铭 李文勇 杨信廷[2] Li Ming;Ding Zhihuan;Zhao Jingxuan;Chen Siming;Li Wenyong;Yang Xinting(School of Computer,Electronics and Information,Guangxi University,Nanning,530004,China;Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences/National Engineering Research Center for Information Technology in Agriculture/National Engineering Laboratory for Agri-product Quality Traceability/Meteorological Service Center for Urban Agriculture,China Meteorological Administration-Ministry of Agriculture and Rural Affairs,Beijing,100097,China)

机构地区：[1]广西大学计算机与电子信息学院,南宁市530004 [2]北京市农林科学院信息技术研究中心/国家农业信息化工程技术研究中心/农产品质量安全追溯技术及应用国家工程研究中心/中国气象局—农业农村部都市农业气象服务中心,北京市100097

出　　处：《中国农机化学报》2023年第5期63-70,共8页Journal of Chinese Agricultural Mechanization

基　　金：国家重点研发计划项目(2022YFE0199500);国家自然科学基金青年科学基金项目(31401683);北京市农林科学院国家基金培育专项(KJCX20211002);北京市农林科学院改革与发展专项——农产品智慧低碳供应链关键技术研究。

摘　　要：针对温室孢子捕捉设备所采集图像中孢子囊分布密集、粘连堆叠和背景复杂的特点,提出一种基于改进YOLOv5s的黄瓜霜霉病孢子囊检测算法。首先,使用带CBAM(Convolutional Block Attention Module)注意力机制的Ghost卷积替代原始网络中的CSP(Cross Stage Partial)模块,抑制背景中的杂质,在保证产生丰富特征图的同时,降低模型的参数量,提升计算速度。其次,修改特征融合网络的连接方式,删除原来负责大物体检测的分支并加入一个更细粒度的分支,以加强对小目标和密集、堆叠目标的检测。最后,对不同预测头产生的损失值赋予不同的权重,并用考虑中心点距离的DIOU_NMS非极大值抑制方法代替原来的NMS方法。改进后的YOLOv5s算法的平均准确率和FPS分别为91.18%和65.4帧/s,比原始的YOLOv5s算法高4.88%和7.1帧/s。该研究可为监测黄瓜霜霉病的发生和发展提供数据支撑,对于保障黄瓜的产量和质量具有重要意义。Aiming at the characteristics of dense sporangia distribution,cohesive stacking and complex background in the images collected by the greenhouse spore capture equipment,an improved YOLOv5s-based sporangia detection algorithm for cucumber downy mildew was proposed.Firstly,a Ghost convolution with CBAM(Convolutional Block Attention Module)attention mechanism was used to replace the CSP(Cross Stage Partial)module in the original network,which suppressed impurities in the background and ensured the generation of rich feature maps while reducing the number of model parameters and improving computational speed.Secondly,the connection method of the feature fusion network was modified,the original branch responsible for large object detection was deleted and a finer-grained branch was added to strengthen the detection of small objects and dense,stacked objects.Finally,different weights were assigned to the loss values generated by different prediction heads,and the original NMS method was replaced by the DIOU_NMS non-maximum value suppression method that considered the center point distance.The average accuracy and FPS of the improved YOLOv5s algorithm were 91.18%and 65.4 frames/s respectively,which were 4.88%and 7.1 frames/s higher than the original YOLOv5s algorithm.This study can provide data support for monitoring the occurrence and development of cucumber downy mildew,and is of great significance for ensuring the yield and quality of cucumbers.

关 键 词：深度学习 目标检测 YOLOv5 黄瓜霜霉病 孢子囊 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]