基于改进MaxViT的辣椒病害识别分类方法
A method for identifying and classifying pepper diseases based on improved MaxViT
李西兴 1陈佳豪 1吴锐 1杨睿1
作者信息
- 1. 湖北工业大学机械工程学院/湖北工业大学现代制造质量工程湖北省重点实验室,武汉 430068
- 折叠
摘要
为实现复杂环境下辣椒病害的精准识别和分类,设计了一种适用于辣椒病害识别分类的方法.以辣椒在生长过程中常见的6种病害为分类研究的对象,使用数据增强的方法扩充数据集,提出一种基于MaxViT改进的MaxViT-DF模型,将MaxViT模型中的普通卷积替换为可变形卷积,使模型在提取特征时能更贴近复杂环境下的识别目标;同时在MaxViT模型施加注意力时引入特征融合模块,提高模型的全局感知能力.结果显示,改进的MaxViT-DF模型识别分类准确率达到 98.10%,对 6种辣椒病害的分类精度均高于 95%.与ResNet-34、EfficientNetv2和VGG-16等模型相比,改进模型在收敛速度和分类精度上具有明显优势.以上结果表明,MaxViT-DF模型能够对不同种类的辣椒常见病害进行有效的分类识别.
Abstract
A method suitable for identifying and classifying pepper diseases was designed to achieve precise recognition and classification of pepper diseases in complex environments including background clut-ter or interference.Six common diseases in the growth process of peppers were used to expand the dataset with data augmentation methods.An improved MaxViT-DF model was proposed based on MaxViT.The or-dinary convolution in the MaxViT model was replaced with deformable convolution to enable the model to extract features closer to the recognition target in complex environments.A feature fusion module was intro-duced when applying attention to the MaxViT model to improve the model's global perception ability.The results showed that the improved MaxViT-DF model had an identification and classification accuracy of 98.10%,and the classification accuracy for six common pepper diseases was higher than 95%.The im-proved model had significant advantages in convergence speed and classification accuracy compared with models including ResNet-34,EfficientNetv2,and VGG-16.It is indicated that the MaxViT-DF model can effectively identify and classify common diseases in different types of peppers.
关键词
MaxViT-DF/辣椒病害分类/可变形卷积/特征融合/深度学习Key words
MaxViT-DF/classification of pepper disease/deformable convolution/feature fusion/deep learning引用本文复制引用
基金项目
国家自然科学基金(51805152)
湖北工业大学绿色工业引领计划(XJ2021005001)
湖北省自然科学基金(2022CFB445)
湖北省重点研发计划(2021BAA203)
出版年
2024
NETL
NSTL
万方数据