河南农业科学 ›› 2022, Vol. 51 ›› Issue (11): 155-162.DOI: 10.15933/j.cnki.1004-3268.2022.11.018

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于改进YOLOX模型的芝麻蒴果检测方法研究

王川1,赵恒滨1,2,李国强2,3,张建涛2,3,高桐梅4,赵巧丽2,3,郑国清2,3   

  1. (1.河南师范大学 计算机与信息工程学院,河南 新乡 453007;2.河南省农业科学院 农业经济与信息研究所/河南省智慧农业工程技术研究中心,河南 郑州 450002;3.农业农村部黄淮海智慧农业技术重点实验室,河南 郑州 450002;4.河南省农业科学院 芝麻研究中心,河南 郑州 450002)
  • 收稿日期:2022-09-22 出版日期:2022-11-15 发布日期:2023-01-10
  • 通讯作者: 李国强(1984-),男,河南林州人,研究员,博士,主要从事农业信息技术研究。E-mail:hnagri@qq.com
  • 作者简介:王川(1976-),男,河南新乡人,副教授,硕士,主要从事人工智能理论及应用研究。E-mail:wangch@htu.edu.cn
  • 基金资助:
    河南省人社厅留学人员科研资助项目;河南省农业科学院农业经济与信息研究所科技创新领军人才培育计划(2021KJCX02);河南省重点研发与推广专项(222102110117)

Detection Method of Sesame Capsules Based on Improved YOLOX Model

WANG Chuan1,ZHAO Hengbin1,2,LI Guoqiang2,3,ZHANG Jiantao2,3,GAO Tongmei4,ZHAO Qiaoli2,3,ZHENG Guoqing2,3   

  1. (1.College of Computer and Information Engineering,Henan Normal University,Xinxiang 453007,China;2.Institute of Agricultural Economics and Information,Henan Academy of Agricultural Sciences/Henan Engineering and Technology Research Center for Intelligent Agriculture,Zhengzhou 450002,China;3.Key Laboratory of Huang‑Huai‑Hai Smart Agricultural Technology,Ministry of Agriculture and Rural Affairs,Zhengzhou 450002,China;4.Sesame Research Center,Henan Academy of Agricultural Sciences,Zhengzhou 450002,China)
  • Received:2022-09-22 Published:2022-11-15 Online:2023-01-10

摘要: 为实现密集条件下芝麻蒴果的准确检测,提出基于YOLOX模型的芝麻蒴果检测定位方法(CEYOLOX模型)。该模型以CSPDarknet-53作为主干特征提取网络,在路径聚合网络PANet中增加104×104大尺度特征层,增强对目标细粒度特征信息的获取;通过引入注意力机制模块获取目标重要的轮廓特征和空间位置信息;将传统的NMS替换为更有利于重叠目标检测的Soft-NMS算法来降低漏检情况。结果表明,在IoU阈值为0.5时,CE-YOLOX模型在全部测试集上的调和均值(F1)、召回率、平均精度分别为0.99、98.65%、99.71%,与原模型YOLOX相比,该模型分别提升了0.05、6.27个百分点、3.28个百分点。通过蒴果计数试验,CE-YOLOX模型计数准确率为96.84%,比YOLOX模型提高了5.28个百分点。改进后的模型CE-YOLOX适用于密集条件下芝麻蒴果检测。

关键词: 芝麻蒴果, 果实检测, 注意力机制, 目标检测算法, YOLOX

Abstract: In order to achieve accurate detection of sesame capsules under dense conditions,this study proposes a sesame capsule detection and localization method based on the YOLOX model(CE‑YOLOX model).In this model,CSPDarknet‑53 is used as the backbone feature extraction network,and a 104×104 large‑scale feature layer is added to the path aggregation network PANet to strengthen the acquisition of the target fine‑grained feature information.By introducing the convolutional block attention module,the important contour features and spatial location information of the object are obtained.The classical NMS is replaced by the Soft‑NMS algorithm,which is more conducive to overlapping target detection,to decrease the missed detection.The results showed that the F1 average of CE‑YOLOX tested on all datasets at IoU threshold of 0.5 was 0.99,0.05 higher than that of YOLOX.The recall rate and average accuracy of CE‑YOLOX were 98.65% and 99.71%,6.27 and 3.28 percentage points higher than that of YOLOX.The counting accuracy of CE‑YOLOX was 96.84%,5.28 percentage points higher than YOLOX.Consequently,the improved model can recognize sesame capsules under dense conditions.

Key words: Sesame capsules, Fruit detection, Attentional mechanism, Target detection algorithm, YOLOX

中图分类号: