基于改进Swin-Unet的小麦条锈病分割方法  

Segmentation Method for Wheat Stripe Rust Based on Improved Swin⁃Unet

作　　者：臧贺藏[1] 任帅 王从胜 王盛威[3] 赵瑞玲 陈丹丹[1] 赵晴[1] 张杰[1] 郑国清[1] 李国强[1] Zang Hecang;Ren Shuai;Wang Congsheng;Wang Shengwei;Zhao Ruiling;Chen Dandan;Zhao Qing;Zhang Jie;Zheng Guoqing;Li Guoqiang(Institute of Agricultural Information Technology,Henan Academy of Agricultural Sciences/Huang-Huai-Hai Key Laboratory of Intelligent Agricultural Technology,Ministry of Agriculture and Rural Affairs,Zhengzhou 450002,China;Xinxiang Academy of Agricultural Sciences,Xinxiang 453600,China;Agricultural Information Institute,Chinese Academy of Agricultural Sciences,Beijing 100081,China;Xinxiang Institue of Engineering,Xinxiang 453706,China)

机构地区：[1]河南省农业科学院农业信息技术研究所/农业农村部黄淮海智慧农业技术重点实验室,河南郑州450002 [2]新乡市农业科学院,河南新乡453600 [3]中国农业科学院农业信息研究所,北京100081 [4]新乡工程学院,河南新乡453706

出　　处：《山东农业科学》2024年第12期147-153,共7页Shandong Agricultural Sciences

基　　金：国家重点研发计划项目(2022YFD2001005);河南省重大科技专项(221100110800);河南省农业科学院自主创新项目(2024ZC069)。

摘　　要：条锈病是影响小麦产量及粮食安全的重要因素,条锈病图像的精准分割是实现计算机辅助精准防治的重要基础。针对小麦条锈病图像中病斑形态复杂、病斑与非病斑之间边界模糊、分割精度低的问题,本研究提出了一种基于改进Swin-Unet的小麦条锈病图像分割方法,通过在Swin-Unet中引入SENet(Squeeze⁃and⁃Excitation Networks)和残差网络(ResNet)模块来增强模型对条锈病特征的表达能力。实验结果表明,改进Swin-Unet对背景、孢子和叶片的查准率分别为99.24%、82.32%和94.36%,可以从复杂环境中有效分割出背景、孢子和叶片图像,具有较好的计算机视觉处理能力和分割评估效果。改进Swin-Unet总体分割准确率、平均交并比和均像素准确率分别为96.88%、84.91%和90.50%,较Swin-Unet分别提高了2.84、4.64个和5.38个百分点;与其他网络模型(U-Net、PSPNet、DeepLabV3+和Swin-Unet)相比,改进Swin-Unet具有最佳分割效果。表明本研究提出的方法可以精准检测和分割小麦条锈病图像,这可为田间复杂环境下小麦条锈病的自动检测和早期预防提供技术支持。Stripe rust is an important factor affecting wheat yield and food security,and accurate seg⁃mentation of wheat stripe rust images is an important means of computer⁃aided precision control.A wheat stripe rust image segmentation method based on improved Swin⁃Unet was proposed in this study to address the prob⁃lems of complex lesion morphology,blurred boundaries between lesions and non⁃lesions,and low segmentation accuracy in wheat stripe rust image.The method enhanced the model’s ability to express stripe rust features by introducing SENet and ResNet modules into Swin⁃Unet.The experimental results showed that the improved Swin⁃Unet had precision rates of 99.24%,82.32%and 94.36%for background,spore and leaf,respectively,and could segment background,spore and leaf images in challenging situations,so it had better computer vi⁃sion processing and segmentation evaluation effects.The overall segmentation accuracy,average intersection to union ratio and average pixel accuracy of improved Swin⁃Unet were 96.88%,84.91%and 90.50%,respective⁃ly,which were 2.84,4.64 and 5.38 percentage points higher than those of Swin⁃Unet.Compared with othernetwork models such as U⁃Net、PSPNet、DeepLabV3+and Swin⁃Unet,the improved Swin⁃Unet had the bestsegmentation performance.The method proposed in this study could accurately detect and segment wheat striperust features,providing technical support for automatic detection and early prevention of wheat stripe rust incomplex field environments.

关 键 词：小麦条锈病 语义分割 Swin-Unet 注意力机制 

分 类 号：S126[农业科学—农业基础科学]