基于Swin-CNN的苹果叶片病斑异常检测研究
Research on apple leaf spot anomaly detection based on Swin-CNN
李文东 1朱幸辉 1邓阳君1
作者信息
- 1. 湖南农业大学 信息与智能科学技术学院,湖南 长沙 410125
- 折叠
摘要
针对当新的病害出现时,可能无法及时检测到的问题,提出了一种应对苹果叶病斑的新方法,即异常检测.采用了一种通用的图像异常增强网络,只需要训练一次,网络(Swin-CNN)就可以对从未见过的病斑叶片进行检测.Swin-CNN在一组无异常图像上进行训练,使用随机掩码,以无监督的方式学习异常与背景之间的空间上下文特征.实验结果表明,Swin-CNN为最优模型,对苹果叶叶斑病、褐斑病、灰斑病、锈病的识别准确率分别达到了88.1%、84.0%、87.0%、97.6%.该方法可以为苹果叶病斑的早期预警提供理论基础,为其他的作物病害检测提供参考.
Abstract
Addressing the issue of potential failure in timely detection when new diseases emerge,a novel approach for dealing with apple leaf spot diseases has been introduced:anomaly detection.A universal image anomaly enhancement network(Swin-CNN)was adopted,which only needs to be trained once to detect diseased leaves that had never been seen before.Swin-CNN was trained on a set of anomaly free images using random masks to learn spatial contextual features between anomalies and background in an unsupervised manner.The experimental results showed that Swin-CNN was the optimal model,achieving an accuracy of 88.1%in identifying apple leaf spot disease,84.0%in identifying apple leaf brown spot disease,87.0%in identifying apple leaf gray spot disease,and 97.6%in identifying apple leaf rust disease.This method can provide a theoretical basis for early warning of apple leaf disease spots and serve as a reference for the detection of other crop diseases.
关键词
Swin-CNN/异常检测/变换域搜索模块/随机掩码/苹果叶病斑Key words
Swin-CNN/anomaly detection/transform domain search module/random masking/apple leaf spots引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据