河南农业科学 ›› 2022, Vol. 51 ›› Issue (6): 154-163.DOI: 10.15933/j.cnki.1004-3268.2022.06.017

所属专题: 作物图像采集与识别专题

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于CSJMM-AS-GAC 的马陆葡萄病虫害识别研究

王兴旺1,郑汉垣2,王素青3   

  1. (1.上海农林职业技术学院,上海 201699;2.上海大学计算机工程与科学学院,上海 200444;3.上海马陆葡萄研究所,上海 201818)
  • 收稿日期:2022-03-03 出版日期:2022-06-15 发布日期:2022-08-17
  • 作者简介:王兴旺(1978-),男,河南开封人,副教授,博士,主要从事果树病虫害防治、设施农业与装备技术方面的研究。 E-mail:gpguiping@163.com
  • 基金资助:
    国家自然科学基金项目(61473237);国家自然科学基金面上项目(61873156)

Identification of Diseases and Insect Pests of Malu Grape Based on CSJMM⁃AS⁃GAC

WANG Xingwang1,ZHENG Hanyuan2,WANG Suqing3   

  1. (1.Shanghai Vocational College of Agriculture and Forestry,Shanghai 201699,China;2.School of Computer Engineering and Science,Shanghai University,Shanghai 200444,China;3.Shanghai Malu Grape Research Institute,Shanghai 201818,China)
  • Received:2022-03-03 Published:2022-06-15 Online:2022-08-17

摘要: 为了提高马陆葡萄病虫害的识别准确率,有效地进行马陆葡萄病虫害防控,对测地线活动轮廓模型(GAC)进行改进,通过引入动态系数函数将马陆葡萄病虫害图像边界区域与非边界区域进行精确划分,从而实现准确分割病虫害图像模糊和凹陷边界,提出并建立了精确分割测地线活动轮廓模型(ASGAC)。接下来为了克服复杂背景下训练样本不足造成的误差,提出了Core损失函数,建立了Core-Softmax联合监督机制(CSJMM),从而确立了基于CSJMM的精确分割测地线活动轮廓模型(CSJMM-ASGAC)。结果表明,CSJMM-AS-GAC训练集初始准确率为65.46%,验证集准确率为95.67%,测试集准确率为93.95%,Kappa系数达到0.913 8,召回率达到89.21%,CSJMM-AS-GAC对于马陆葡萄病虫害识别准确率达到94.06%。CSJMM-AS-GAC的整体性能、识别准确率、召回率等指标都优于常用的病虫害识别模型。

关键词: 马陆葡萄, 病虫害识别, 损失函数, 分割, 测地线活动轮廓模型, 精确分割测地线活动轮廓模型, 召回率

Abstract: In order to improve the identification accuracy of diseases and insect pests of Malu grape and effectively carry out the prevention and control of diseases and insect pests of Malu grape,this paper improved the geodesic active contour model(GAC),by introducing the dynamic coefficient function,the boundary region and non⁃boundary region of Malu grape pest image were accurately divided,so as to accurately segment the fuzzy and concave boundary of the pest image. A precise segmented geodesic active contour model(AS⁃GAC) was proposed and established. Next,in order to overcome the error caused by insufficient training samples in complex background,the Core loss function was proposed,and the Core⁃Softmax joint supervision mechanism(CSJMM) was established. Thus,a precise segmented geodesic active contour model based on CSJMM(CSJMM⁃AS⁃GAC)was established. The results showed that the initial accuracy of CSJMM⁃AS⁃GAC training set was 65.46%,the accuracy of validation set was 95.67%,the accuracy of test set was 93.95%,the Kappa coefficient of CSJMM⁃AS⁃GAC reached 0.913 8,the recall rate reached 89.21%,and the average accuracy of CSJMM⁃AS⁃GAC for identifying the diseases and pests of Malu grape reached 94. 06%. The overall performance,recognition accuracy and recall rate of CSJMM⁃AS⁃GAC are better than the commonly used disease and pest identification models.

Key words: Malu grape, Disease and pest identification, Loss function, Division, GAC, AS?GAC, Recall rate

中图分类号: