基于多尺度注意力增强Res2Net50的番茄病虫害图像分类方法

李海棠

绿洲农业科学与工程 ›› 2026, Vol. 11 ›› Issue (03) : 85 -90.

PDF (1169KB)
绿洲农业科学与工程 ›› 2026, Vol. 11 ›› Issue (03) : 85 -90. DOI: 10.26941/j.cnki.2096-2177.2026.03.013
植物保护·微生物

基于多尺度注意力增强Res2Net50的番茄病虫害图像分类方法

作者信息 +

Tomato Pest and Disease Image Classification Method Based on Multi-Scale Attention-Enhanced Res2Net50

Author information +
文章历史 +
PDF (1197K)

摘要

针对番茄病害识别任务中存在的多尺度特征表达不足、细微病斑难以捕捉以及复杂背景干扰明显等问题,本文提出一种基于多尺度注意力机制的改进残差分辨率网络50层(Res2Net50)模型。该方法在Res2Net50的瓶颈二连块(Bottle2neck)模块中引入一种新颖的多尺度注意力机制(Multi-Scale Attention Mechanism,MSAM),通过尺度内通道注意力强化各尺度特征的语义表达能力,并利用跨尺度交互注意力建模尺度间的依赖关系,从而实现对病害多尺度特征的差异化加权与自适应融合。该机制在几乎不增加计算负担的前提下,显著提升模型对多尺度病斑特征的感知与区分能力。在公开的番茄叶片病害数据集上的试验结果表明,改进模型取得96.54%的识别准确率,优于原始Res2Net50(94.82%)、密集连接网络121层(DenseNet121,94.29%)、轻量移动网络V2(MobileNetV2,90.64%)及混洗网络V2(ShuffleNetV2,82.64%)等主流模型。消融试验进一步证实多尺度注意力模块的有效性,其引入使模型准确率相对提升1.72%。本文所提方法不仅显著提升番茄病害识别的精度与鲁棒性,也为复杂农业场景下的作物病害智能诊断提供可借鉴的技术选择。

Abstract

To address the issues of insufficient multi-scale feature representation,difficulty in capturing subtle disease spots,and significant interference from complex backgrounds in tomato disease recognition tasks,this paper proposes an improved Res2Net50 model based on a multi-scale attention mechanism.The method introduces a novel Multi-Scale Attention Mechanism(MSAM)into the Bottle2neck module of Res2Net50.It enhances the semantic representation capability of features at each scale through intra-scale channel attention and establishes dependencies between scales via cross-scale interactive attention,thereby achieving differentiated weighting and adaptive fusion of multi-scale disease features.This mechanism significantly improves the model's perception and discrimination ability for multi-scale disease features with minimal computational overhead.Experimental results on a public tomato leaf disease dataset show that the improved model achieves a recognition accuracy of 96.54%,outperforming mainstream models such as the original Res2Net50(94.82%),Dense connection network with 121 layers(DenseNet121 94.29%),Lighweight Mobile Network V2(MobileNetV2 90.64%),and Mixed-washing Network V2(ShuffleNetV2 82.64%).Ablation experiments further verify the effectiveness of the multi-scale attention module,whose introduction improves the model accuracy by 1.72%.The proposed method not only significantly enhances the accuracy and robustness of tomato disease recognition,but also provides a referable technical option for intelligent crop disease diagnosis in complex agricultural scenarios.

关键词

番茄病害识别 / Res2Net50 / 多尺度注意力机制 / 特征融合 / 深度学习

Key words

tomato disease recognition / Res2Net50 / multi-scale attention mechanism / feature fusion / deep learning

引用本文

引用格式 ▾
李海棠. 基于多尺度注意力增强Res2Net50的番茄病虫害图像分类方法[J]. 绿洲农业科学与工程, 2026, 11(03): 85-90 DOI:10.26941/j.cnki.2096-2177.2026.03.013

登录浏览全文

4963

注册一个新账户 忘记密码

参考文献

[1]

倪智涛, 胡伟健, 李宝山, . 一种基于图像分类与目标检测协同的番茄细粒度病害识别方法[J]. 江苏农业科学202351(22):221-228.

[2]

NI ZhitaoHU WeijianLI Baoshan ,et al. A tomato fine—grained disease recognition method based on the collaboration of image classification and object detection[J]. Jiangsu Agricultural Sciences202351(22):221-228.

[3]

孔祥源, 王一群, 缪祎晟, . 基于TomatoVit的番茄病害分类与分级研究[J]. 重庆理工大学学报(自然科学)202539(9):133-141.

[4]

KONG XiangyuanWANG YiqunMIAO Yisheng ,et al. Research on tomato disease classification and grading based on TomatoVit[J]. Journal of Chongqing University of Technology(Natural Science)202539(9):133-141.

[5]

姜晟久, 钟国韵. 基于可分离扩张卷积和通道剪枝的番茄病害分类方法[J]. 江苏农业科学202452(2):182-189.

[6]

JIANG ShengjiuZHONG Guoyun . A tomato disease classification method based on separable dilated convolution and channel pruning[J]. Jiangsu Agricultural Sciences202452(2):182-189.

[7]

马丽, 周巧黎, 赵丽亚, . 基于深度学习的番茄叶片病害分类识别研究[J]. 中国农机化学报202344(7):187-193+206.

[8]

MA LiZHOU QiaoliZHAO Liya ,et al. Research on tomato leaf disease classification and recognition based on deep learning[J]. Journal of Chinese Agricultural Mechanization202344(7):187-193+206.

[9]

刘拥民, 刘翰林, 石婷婷, . 一种优化的Swin Transformer番茄叶片病害识别方法[J]. 中国农业大学学报202328(4):80-90.

[10]

LIU YongminLIU HanlinSHI Tingting ,et al. An optimized Swin Transformer method for tomato leaf disease recognition[J]. Journal of China Agricultural University202328(4):80-90.

[11]

汤文亮, 黄梓锋. 基于知识蒸馏的轻量级番茄叶部病害识别模型[J]. 江苏农业学报202137(3):570-578.

[12]

TANG WenliangHUANG Zifeng . A lightweight tomato leaf disease recognition model based on knowledge distillation[J]. Jiangsu Journal of Agricultural Sciences202137(3):570-578.

[13]

Gao S HCheng M MZhao K ,et al. Res2Net:A new multi—scale backbone architecture[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence202143(2):652-662.

[14]

Ran HWen SWang S ,et al. Memristor—based edge computing of ShuffleNetV2 for image classification[J]. IEEE Transactions on Computer—Aided Design of Integrated Circuits and Systems202140(8):1 701—1 710.

[15]

Sandler MHoward A GZhu M ,et al. MobileNetV2:inverted residuals and linear bottlenecks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Salt Lake City:IEEE, 2018:4 510—4 520.

[16]

Huang GLiu ZVan DMaaten L ,et al. Densely connected convolutional networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.Honolulu:IEEE, 2017:2 261—2 269.

基金资助

基于计算机视觉的周口市重要场所安全防范系统优化研究(ZKSKDY-2025-685)

AI Summary AI Mindmap
PDF (1169KB)

0

访问

0

被引

详细

导航
相关文章

AI思维导图

/