[1]苏航,陈旭昊,寿德荣,等.基于注意力机制轻量化模型的植物病害识别方法[J].江苏农业学报,2024,(08):1389-1399.[doi:doi:10.3969/j.issn.1000-4440.2024.08.004]
　SU Hang,CHEN Xuhao,SHOU Derong,et al.Plant disease recognition method based on lightweight model with attention mechanism[J].,2024,(08):1389-1399.[doi:doi:10.3969/j.issn.1000-4440.2024.08.004]
点击复制

基于注意力机制轻量化模型的植物病害识别方法()
分享到：

江苏农业学报[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2024年08期
页码:
1389-1399
栏目:
植物保护
出版日期:
2024-08-30

文章信息/Info

Title:
Plant disease recognition method based on lightweight model with attention mechanism
作者:
苏航1; 2; 3; 陈旭昊1; 2; 寿德荣1; 张朝阳1; 许彪3; 孙丙宇2
(1.重庆三峡学院机械工程学院,重庆404100;2.中国科学院合肥物质科学研究院智能机械研究所,安徽合肥230000;3.中国工程物理研究院,四川成都610000)
Author(s):
SU Hang1; 2; 3; CHEN Xuhao1; 2; SHOU Derong1; ZHANG Chaoyang1; XU Biao3; SUN Bingyu2
(1.School of Mechanical Engineering, Chongqing Three Gorges University, Chongqing 404100, China;2.Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences, Hefei 230000, China;3.China Academy of Engineering Physics, Chengdu 610000, China)
关键词:
模型剪枝; 卷积块注意力模块(CBAM)注意力机制; 大卷积核倒置残差结构（IRBCKS）模块; 植物病害; 轻量化网络
Keywords:
model pruning; convolutional block attention module (CBAM) attention mechanism; inverted residual block convolution kernel structure (IRBCKS) module; plant diseases; lightweight networks
分类号:
S608
DOI:
doi:10.3969/j.issn.1000-4440.2024.08.004
文献标志码:
A
摘要:
针对现有植物病害识别模型存在响应速度慢、参数量多、计算机内存资源消耗大等问题,本研究提出了一种轻量化神经网络模型,该模型由特征提取层、特征增强层和分类器组成。为了减小模型大小并提高网络响应速度,在特征提取层中使用深度可分离卷积进行特征提取。为了防止网络传播过程中的梯度消失并增强病害像素特征融合,在特征提取层中引入了大卷积核倒置残差结构（IRBCKS）模块。此外,在特征增强层集成了轻量级卷积块注意力模块(CBAM)注意力机制,以捕捉植物病害相关图像中像素之间的关系,增强关键信息的提取。最后,采用剪枝技术剔除模型中冗余特征信息,从而再次减少模型参数量,形成最终的轻量级网络模型Cut-MobileNet。为验证该模型的先进性,将其与轻量化模型（MobileNet V2、SqueezeNet、GoogLeNet）和非轻量化模型（Vision Transformer、AlexNet）进行性能对比,研究结果表明,Cut-MobileNet在浮点运算量、准确率、单张图片推理时间、参数量、F1值和模型大小等性能指标上都取得了较优的效果。
Abstract:
In light of the issues associated with slow response speed, numerous parameters, and high computational memory requirements in existing plant disease recognition models, we proposed a lightweight neural network model. The model consisted of feature extraction layer, feature enhancement layer, and classifier. To reduce model size and increase network response speed, we utilized deep separable convolution in the feature extraction layer. To prevent gradient disappearance during network propagation and enhance the fusion of disease pixel features, we introduced the inverted residual block convolution kernel structure (IRBCKS) module into the feature extraction layer. Furthermore, we integrated a lightweight convolutional block attention module (CBAM) attention mechanism into the feature enhancement layer to capture the relationships between pixels in plant disease-related images and enhance key information extraction. Finally, we employed a pruning technique to eliminate redundant feature information from the base model, thereby reducing the number of model parameters once again, yielding this lightweight network model, Cut-MobileNet. In order to verify the progressiveness of this model, it was compared with lightweight models (MobileNet V2, SqueezeNet, GoogLeNet) and non-lightweigh models (Vision Transformer, AlexNet). The results show that better results have been achieved by Cut-MobileNet in floating-point operation, accuracy, single image inference time, parameter count, F1 value, and model size.

参考文献/References:

[1]TIAN H K, WANG T H, LIU Y D, et al. Computer vision technology in agricultural automation —a review[J]. Information Processing in Agriculture,2020,7(1):1-19.
[2]刘拥民,刘翰林,石婷婷,等. 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,2023,28(4):80-90.
[3]马丽,周巧黎,赵丽亚,等. 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,2023,44(7):187-193.

[4]牛学德,高丙朋,南新元,等. 基于改进DenseNet卷积神经网络的番茄叶片病害检测[J]. 江苏农业学报,2022,38(1):129-134.
[5]YAZDIAN H, SALMANI-DEHAGHI N, ALIJANIAN M. A spatially promoted SVM model for GRACE downscaling: using ground and satellite-based datasets[J]. Journal of Hydrology,2023,626:130214.
[6]KHALIFA F, ABDELKADER H, ELSAID A. An analysis of ensemble pruning methods under the explanation of Random Forest[J]. Information Systems,2024,120:102310.
[7]GUAN X, TERADA Y. Sparse kernel K-means for high-dimensional data[J]. Pattern Recognition,2023,144:109873.
[8]崔兆亿,耿秀丽. 基于随机森林和量子粒子群优化的SVM算法[J]. 计算机集成制造系统,2023,29(9):2929-2936.
[9]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,2021,37(3):570-578.
[10]ABADE A, FERREIRA P, DE BARROS VIDAL F. Plant diseases recognition on images using convolutional neural networks: a systematic review[J]. Computers and Electronics in Agriculture,2021,185:106125.
[11]SMIRNOV E, TIMOSHENKO D, ANDRIANOV S. Comparison of regularization methods for imageNet classification with deep convolutional neural networks[J]. AASRI Procedia,2014,6:89-94.
[12]JIA S J, JIA P Y, HU S P, et al. Automatic detection of tomato diseases and pests based on leaf images:2017 Chinese automation congress (CAC)[C]. Jinan:IEEE, 2017.
[13]YADAV D, JALAL A, GARLAPATI D, et al. Deep learning-based ResNeXt model in phycological studies for future[J]. Algal Research,2020,50:102018.
[14]KHAN M, UDDIN M, PARVEZ M, et al. A squeeze and excitation ResNeXt-based deep learning model for Bangla handwritten compound character recognition[J]. Journal of King Saud University-Computer and Information Sciences,2022,34(6):3356-3364.
[15]YANG L, YU X Y, ZHANG S P, et al. GoogLeNet based on residual network and attention mechanism identification of rice leaf diseases[J]. Computers and Electronics in Agriculture,2023,204:107543.
[16]SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]//IEEE. 2015 IEEE Conference on Computer Vision and Pattern Recognition(CVPR). Boston,MA,USA,2015:1-9.
[17]RAZA M, NOSHEEN A, YASMIN H, et al. Application of aquatic plants alone as well as in combination for phytoremediation of household and industrial wastewater[J]. Journal of King Saud University-Science,2023,35(7):102805.
[18]LU Q, YE W X, YIN L F. ResDenIncepNet-CBAM with principal component analysis for wind turbine blade cracking fault prediction with only short time scale SCADA data[J]. Measurement,2023,212:112696.
[19]CHEN L J, YAO H D, FU J Y, et al. The classification and localization of crack using lightweight convolutional neural network with CBAM[J]. Engineering Structures,2023,275:115291.
[20]LAU K, PO L, REHMAN Y. Large separable kernel attention: rethinking the large kernel attention design in CNN[J]. Expert Systems with Applications,2023,236:121352.
[21]杨佳昊,左昊轩,黄祺成,等. 基于YOLO v5s的作物叶片病害检测模型轻量化方法[J]. 农业机械学报,2023,54(增刊1):222-229.
[22]陈智超,汪国强,李飞,等. 基于Bi-LSTM与多尺度神经网络模型的番茄病害识别[J]. 江苏农业科学,2023,51(15):194-203.
[23]曹林,周凯,申鑫,等. 智慧林业发展现状与展望[J]. 南京林业大学学报(自然科学版),2022,46(6):83-95.
[24]张会敏,谢泽奇. 基于知识图谱与深度学习的黄瓜叶部病害识别方法[J]. 江苏农业科学, 2023,51(15):173-178.
[25]鲍彤,罗瑞,郭婷,等. 基于BERT字向量和TextCNN的农业问句分类模型分析[J]. 南方农业学报,2022,53(7):2068-2076.
[26]刘媛媛,王定坤,邬雷,等. 基于知识蒸馏和模型剪枝的轻量化模型植物病害识别[J]. 浙江农业学报,2023,35(9):2250-2264.
[27]邵仁荣,刘宇昂,张伟,等. 深度学习中知识蒸馏研究综述[J]. 计算机学报,2022,45(8):1638-1673.

备注/Memo

备注/Memo:
收稿日期：2023-11-29基金项目：国家自然科学基金项目(61773360);2019年重庆市人工智能+智慧农业学科群开放基金项目(ZNNYKFA201901)作者简介：苏航(1996-),男,陕西渭南人,硕士研究生,研究方向为智能制造、农业信息化研究、图像处理等。(E-mail)935659856@qq.com通讯作者：孙丙宇,(E-mail)bysun@iim.ac.cn
更新日期/Last Update: 2024-09-18