摘要
传统马铃薯晚疫病检测方法在捕捉微小病斑和准确定位病害边缘时往往存在局限性,导致分类模型对病斑的识别精度有限。提出了一种基于双重注意力网络的马铃薯晚疫病分类模型。该模型以2D、3D结合的混合卷积神经网络为主干网络,结合3D并行空间和通道压缩激励(3D concurrent spatial and channel squeeze&excitation,3DscSE)模块,从通道和空间2个维度对特征进行加权,以增强模型对重要特征的关注能力。为增强边缘像素与中心像素的关系,融入十字交叉注意力(criss-cross attention,CCA)机制捕捉微小的病变区域。实验结果表明,该模型在马铃薯晚疫病叶片检测上能够捕捉病变区域的边界和细节,模型分类总体精度为91.16%。
The traditional detection methods for potato late blight often have limitations in capturing small disease spots and accurately locating disease edges,resulting in limited recognition accuracy of disease spots by classification models.A potato late blight classification model based on a dual attention network was proposed.This model uses a hybrid convolutional neural network that combines 2D and 3D as the backbone network,combined with a 3D concurrent spatial and channel squeeze&excitation(3DscSE)module,which weights features from both channel and spatial dimensions to enhance the model's ability to focus on important features.To enhance the relationship between edge pixels and center pixels,a criss-cross attention(CCA)mechanism was incorporated to capture small lesion areas.The experimental results show that the model can capture the boundaries and details of the lesion area in the detection of potato late blight leaves,and the overall classification accuracy of the model is 91.16%.
作者
刘雨琛
张巧杰
LIU Yuchen;ZHANG Qiaojie(College of Automation,Beijing Information Science&Technology University,Beijing 100192,China)
出处
《北京信息科技大学学报(自然科学版)》
2025年第3期29-35,共7页
Journal of Beijing Information Science and Technology University(Science and Technology Edition)
基金
国家重点研发计划项目(2021ZD0113603)。
关键词
高光谱图像
马铃薯晚疫病
深度学习
注意力机制
hyperspectral images
potato late blight
deep learning
attention mechanism
