Journal of Integrative Agriculture ›› 2024, Vol. 23 ›› Issue (10): 3554-3575.DOI: 10.1016/j.jia.2023.11.037

• • 上一篇    下一篇

基于KVA和GR-ARNet的香蕉叶病识别

  

  • 收稿日期:2023-04-03 接受日期:2023-11-07 出版日期:2024-10-20 发布日期:2024-09-11

Identification of banana leaf disease based on KVA and GR-ARNet

Jinsheng Deng1*, Weiqi Huang1*, Guoxiong Zhou1#, Yahui Hu2, Liujun Li3, Yanfeng Wang4   

  1. 1 College of Electronic Information and Physics, Central South University of Forestry and Technology, Changsha 410004, China
    2 Plant Protection Research Institute, Hunan Academy of Agricultural Sciences, Changsha 410125, China
    3 Department of Soil and Water Systems, College of Agricultural & Life Sciences, University of Idaho, Moscow 83844, USA
    4 College of Systems Engineering, National University of Defense Technology, Changsha 410073, China

  • Received:2023-04-03 Accepted:2023-11-07 Online:2024-10-20 Published:2024-09-11
  • About author:Jinsheng Deng, E-mail: dengjs2023@163.com; Weiqi Huang, E-mail: weiqi_237@163.com; #Correspondence Guoxiong Zhou, E-mail: zhougx01@163.com * These authors contributed equally to this study.
  • Supported by:
    This work was supported by the Changsha Municipal Natural Science Foundation, China (kq2014160); in part by the Key Projects of Department of Education of Hunan Province, China (21A0179), the Hunan Key Laboratory of Intelligent Logistics Technology, China (2019TP1015)  and the National Natural Science Foundation of China (61902436).

PDF

摘要：

据世界粮食及农业组织统计，香蕉是世界第二大水果作物，也是贸易量和消费量最大的水果之一，它们在热带和亚热带国家被广泛种植。香蕉叶斑病、香蕉灰纹病和香蕉拟盘多毛菌病等香蕉叶病有可能对香蕉生产造成严重影响。在香蕉叶病检测中，存在着香蕉叶图像噪声干扰和蕉叶疾病具有相似性等问题的干扰，利用机器视觉和神经网络识别香蕉叶病仍然具有挑战性。针对上述问题，本文提出了一种新的方法来识别香蕉叶病。首先，提出一种名为K-scale VisuShrinkd algorithm (KVA)的新型算法对香蕉叶片图像进行去噪，该算法在半软阈值和中程阈值的基础上引入新的分解尺度k，获得理想的阈值解，并用新建立的阈值函数进行替代，从而达到图像降噪的效果，得到噪声较小的香蕉叶片图像。然后，本文在Resnet50网络架构的基础上提出了一种香蕉叶病识别的新型网络，称为 Ghost ResNeSt-Attention RReLU-Swish Net (GR-ARNet)。其中，引入Ghost模块处理蕉叶病信息的冗余特征图，有利于网络对输入特征图信息全面理解，提高网络提取蕉叶病害深度特征信息的效率和识别速度；采用ResNeSt模块调整各通道的权重，增强蕉叶病有用特征信息的通道，抑制注意学习中噪声信息的通道，增强网络对深度特征的识别能力，提高蕉叶病特征提取能力，从而获取到详细的蕉叶病斑特征图，降低相似病害识别的错误率；利用RReLU和Swish的混合激活函数加快模型的训练速度，提高网络的泛化能力。本文提出的模型对13021张香蕉叶病图像的平均准确率为96.98%，精确率为89.31%，每秒内可以处理的香蕉叶图像为83张。实验结果表明，本文提出的模型具有较高的识别精度和识别速度，在农业病害防治中具有重要的应用价值。

Abstract:

Banana is a significant crop, and three banana leaf diseases, including Sigatoka, Cordana and Pestalotiopsis, have the potential to have a serious impact on banana production.  Existing studies are insufficient to provide a reliable method for accurately identifying banana leaf diseases.  Therefore, this paper proposes a novel method to identify banana leaf diseases.  First, a new algorithm called K-scale VisuShrink algorithm (KVA) is proposed to denoise banana leaf images.  The proposed algorithm introduces a new decomposition scale K based on the semi-soft and middle course thresholds, the ideal threshold solution is obtained and substituted with the newly established threshold function to obtain a less noisy banana leaf image.  Then, this paper proposes a novel network for image identification called Ghost ResNeSt-Attention RReLU-Swish Net (GR-ARNet) based on Resnet50.  In this, the Ghost Module is implemented to improve the network’s effectiveness in extracting deep feature information on banana leaf diseases and the identification speed; the ResNeSt Module adjusts the weight of each channel, increasing the ability of banana disease feature extraction and effectively reducing the error rate of similar disease identification; the model’s computational speed is increased using the hybrid activation function of RReLU and Swish.  Our model achieves an average accuracy of 96.98% and a precision of 89.31% applied to 13,021 images, demonstrating that the proposed method can effectively identify banana leaf diseases.


Key words: banana leaf diseases ,  image denoising ,  Ghost Module ,  ResNeSt Module ,  Convolutional Neural Networks ,  GR-ARNet