基于迁移改进MobileNetV3的玉米叶片病害识别研究

doi:10.11924/j.issn.1000-6850.casb2024-0773

摘要/Abstract

摘要：

近年来，深度学习算法在图像识别领域的应用已逐渐拓展到农业生产中，特别是在作物病害检测方面。结合深度学习中的迁移学习技术，笔者提出了一种基于MobileNetV3改进模型的玉米叶片病害识别方法。将在ImageNet数据集上预训练的权重迁移到目标数据集，并在此基础上对模型进行优化。优化过程中采用CBAM注意力模块替换了原有的SE模块，并在卷积层中引入空洞卷积以增大感受野。经过训练得到了一个最优玉米叶片病害识别模型。经过迁移学习后，模型在训练集上的准确率由96.30%提升至98.20%，提高了1.9个百分点。在此基础上进一步优化后，模型准确率达到99.09%，识别效果更为优异。该改进不仅保留了MobileNetV3的轻量化特性，而且显著提高了玉米叶片病害识别的性能。

关键词: 玉米病害识别, 迁移学习, MobileNetV3, 注意力机制

Abstract:

In recent years, the application of deep learning algorithms in the field of image recognition has gradually expanded into agricultural production, particularly in the area of crop disease detection. Leveraging transfer learning techniques within deep learning, a method for identifying corn leaf diseases based on an improved MobileNetV3 model has been proposed. Pre-trained weights from the ImageNet dataset were transferred to the target dataset, and the model was further optimized. During the optimization process, the original SE (Squeeze-and-Excitation) module was replaced with a CBAM (Convolutional Block Attention Module) attention module, and dilated convolutions were introduced into the convolutional layers to increase the receptive field. After training, an optimal model for corn leaf disease identification was obtained. After applying transfer learning, the model's accuracy on the training set increased from 96.30% to 98.20%, with an improvement of 1.9 percentage points. With further optimization, the accuracy reached 99.09%, demonstrating improved classification performance. This enhancement not only retains the lightweight characteristics of MobileNetV3 but also significantly boosts the performance of corn leaf disease identification.

Key words: corn disease identification, transfer learning, MobileNetV3, attention mechanism

金雨纯, 甄元元, 刘平. 基于迁移改进MobileNetV3的玉米叶片病害识别研究[J]. 中国农学通报, 2025, 41(23): 145-154.

JIN Yuchun, ZHEN Yuanyuan, LIU Ping. Maize Leaf Disease Recognition Based on Transferred and Improved MobileNetV3[J]. Chinese Agricultural Science Bulletin, 2025, 41(23): 145-154.

图/表 16

参考文献 27

[1]	李少昆, 王崇桃. 我国玉米产量变化及增产因素分析[J]. 玉米科学, 2008(4):26-30.
[2]	张娣. 我国玉米病害现状及防治措施[J]. 南方农机, 2017, 48(17):57.
[3]	MILLER S A, BEED F D, HARMON C L. Plant disease diagnostic capabilities and networks[J]. Annual review of phytopathology, 2009, 47(1):15-38.
[4]	佘奎军, 刘永进, 贺奇, 等. 2001—2015年宁夏玉米病虫害发生特征与损失分析[J]. 环境昆虫学报, 2017, 39(5):1071-1080.
[5]	WANG H, LIN Y, XU X, et al. A study on long-close distance coordination control strategy for litchi picking[J]. Agronomy, 2022, 12(7):1520.
[6]	WEN C, WU J, CHEN H, et al. Wheat spike detection and counting in the field based on SpikeRetinaNet[J]. Frontiers in plant science, 2022, 13:821717.
[7]	KANG J, LIU L, ZHANG F, et al. Semantic segmentation model of cotton roots in-situ image based on attention mechanism[J]. Computers and electronics in agriculture, 2021, 189:106370.
[8]	江顺, 黄红星, 莫里楠, 等. 基于改进AlexNet的岭南水稻虫害识别方法研究[J]. 江苏农业科学, 2023, 51(23):187-195.
[9]	李明泽, 张静, 雷亚平, 等. 基于卷积神经网络的棉花关键生育时期识别[J]. 核农学报, 2024, 38(10):2020-2031. doi: 10.11869/j.issn.1000-8551.2024.10.2020
[10]	de SÁ LEITÃO D A H, SHARMA A K, SINGH A, et al. Yield and plant height predictions of irrigated maize through unmanned aerial vehicle in North Florida[J]. Computers and electronics in agriculture, 2023, 215:108374.
[11]	王大庆, 禄琳, 于兴龙, 等. 基于深度迁移学习的EfficientNet玉米叶部病害识别[J]. 东北农业大学学报, 2023, 54(5):66-76.
[12]	杨明欣, 张耀光, 刘涛. 基于卷积神经网络的玉米病害小样本识别研究[J]. 中国生态农业学报(中英文), 2020, 28(12):1924-1931.
[13]	凌慕菲, 杨冬风. 基于改进YOLOv5s模型的玉米叶片病害识别[J]. 农业灾害研究, 2023, 13(8):126-128.
[14]	CHENG H B, LI H M. Identification of apple leaf disease via novel attention mechanism based convolutional neural network[J]. Frontiers in plant science, 2023, 14:1274231.
[15]	XIE W, WEI S, ZHENG Z, et al. Recognition of defective carrots based on deep learning and transfer learning[J]. Food and bioprocess technology, 2021, 14(7):1361-1374.
[16]	崔金荣, 魏文钊, 赵敏. 基于改进MobileNetV3的水稻病害识别模型[J]. 农业机械学报, 2023, 54(11):217-224.
[17]	SI H, WANG Y, ZHAO W, et al. Apple surface defect detection method based on weight comparison transfer learning with MobileNetV3[J]. Agriculture, 2023, 13(4):824.
[18]	师翊, 王应宽, 王菲, 等. MCBPnet:一种高效的轻量级青杏识别模型[J]. 农业工程学报, 2025, 41(5):156-164.
[19]	HAN D, LIU Q, FAN W. A new image classification method using CNN transfer learning and web data augmentation[J]. Expert systems with applications, 2018, 95:43-56.
[20]	吴迪, 肖衍, 沈学军, 等. 基于改进Res2Net与迁移学习的水果图像分类[J]. 电子科技大学学报, 2025, 54(1):62-71.
[21]	刘鑫鹏, 栾悉道, 谢毓湘, 等. 迁移学习研究和算法综述[J]. 长沙大学学报, 2018, 32(5):28-31.
[22]	王中璞, 吴正香, 尤美路, 等. 基于改进MobileNetV2的棉花颜色分级检测[J]. 棉纺织技术, 2024, 52(6):15-21.
[23]	戴敏, 孙文靖, 缪宏. 基于轻量化CBAM-GoogLeNet的辣椒病虫害识别[J]. 中国农机化学报, 2025, 46(2):224-229.
[24]	谭泗桥, 陈涵, 朱磊, 等. 基于改进YOLOv8m的稻田害虫识别方法[J]. 农业工程学报, 2025, 41(2):185-195.
[25]	徐胜军, 欧阳朴衍, 郭学源, 等. 多尺度特征融合空洞卷积ResNet遥感图像建筑物分割[J]. 光学精密工程, 2020, 28(7):1588-1599.
[26]	DOU S, WANG L, FAN D, et al. Classification of Citrus Huanglongbing Degree based on CBAM-MobileNetV2 and transfer learning[J]. Sensors, 2023, 23(12):5587.
[27]	赖勤波, 马正华, 朱蓉. 基于注意力机制和空洞卷积的无人机图像目标检测[J]. 计算机应用与软件, 2025, 42(2):227-235.

输入尺寸（高×宽×通道数）	卷积操作	扩展尺寸	输出通道	SE模块	激活函数	步长
224×224×3	conv2d	-	16	-	HS	2
112×112×16	bneck，3×3	16	16	-	RE	1
112×112×16	bneck，3×3	64	24	-	RE	2
56×56×24	bneck，3×3	72	24	-	RE	1
56×56×24	bneck，5×5	72	40	√	RE	2
28×28×40	bneck，5×5	120	40	√	RE	1
28×28×40	bneck，5×5	120	40	√	RE	1
28×28×40	bneck，3×3	240	80	-	HS	2
14×14×80	bneck，3×3	200	80	-	HS	1
14×14×80	bneck，3×3	184	80	-	HS	1
14×14×80	bneck，3×3	184	80	-	HS	1
14×14×80	bneck，3×3	480	112	√	HS	1
14×14×112	bneck，3×3	672	112	√	HS	1
14×14×112	bneck，5×5	672	160	√	HS	2
7×7×160	bneck，5×5	960	160	√	HS	1
7×7×160	bneck，5×5	960	160	√	HS	1
7×7×160	conv2d，1×1	-	900	-	HS	1
7×7×960	pool，7×7	-	-	-	-	1
1×1×960	conv2d，1×1，NBN	-	1280	-	HS	1
1×1×1280	conv2d，1×1，NBN	-	k	-	-	1

输入尺寸（高×宽×通道数）	卷积操作	扩展尺寸	输出通道	SE模块	激活函数	步长
224×224×3	conv2d	-	16	-	HS	2
112×112×16	bneck，3×3	16	16	-	RE	1
112×112×16	bneck，3×3	64	24	-	RE	2
56×56×24	bneck，3×3	72	24	-	RE	1
56×56×24	bneck，5×5	72	40	√	RE	2
28×28×40	bneck，5×5	120	40	√	RE	1
28×28×40	bneck，5×5	120	40	√	RE	1
28×28×40	bneck，3×3	240	80	-	HS	2
14×14×80	bneck，3×3	200	80	-	HS	1
14×14×80	bneck，3×3	184	80	-	HS	1
14×14×80	bneck，3×3	184	80	-	HS	1
14×14×80	bneck，3×3	480	112	√	HS	1
14×14×112	bneck，3×3	672	112	√	HS	1
14×14×112	bneck，5×5	672	160	√	HS	2
7×7×160	bneck，5×5	960	160	√	HS	1
7×7×160	bneck，5×5	960	160	√	HS	1
7×7×160	conv2d，1×1	-	900	-	HS	1
7×7×960	pool，7×7	-	-	-	-	1
1×1×960	conv2d，1×1，NBN	-	1280	-	HS	1
1×1×1280	conv2d，1×1，NBN	-	k	-	-	1

输入尺寸（高×宽×通道数）	卷积操作	扩展尺寸	输出通道	SE模块	激活函数	步长
224×224×3	conv2d，3×3	-	16	-	HS	2
112×112×16	bneck，3×3	16	16	√	RE	2
56×56×16	bneck，3×3	72	24	-	RE	2
28×28×24	bneck，3×3	88	24	-	RE	1
28×28×24	bneck，5×5	96	40	√	HS	2
14×14×40	bneck，5×5	240	40	√	HS	1
14×14×40	bneck，5×5	240	40	√	HS	1
14×14×40	bneck，5×5	120	48	√	HS	1
14×14×48	bneck，5×5	144	48	√	HS	1
14×14×48	bneck，5×5	288	96	√	HS	2
7×7×96	bneck，5×5	576	96	√	HS	1
7×7×96	bneck，5×5	576	96	√	HS	1
7×7×96	conv2d，1×1	-	576	√	HS	1
7×7×576	pool，7×7	-	-	-	-	1
1×1×576	conv2d，1×1，NBN	-	1024	-	HS	1
1×1×1024	conv2d，1×1，NBN	-	k	-	-	1

输入尺寸（高×宽×通道数）	卷积操作	扩展尺寸	输出通道	SE模块	激活函数	步长
224×224×3	conv2d，3×3	-	16	-	HS	2
112×112×16	bneck，3×3	16	16	√	RE	2
56×56×16	bneck，3×3	72	24	-	RE	2
28×28×24	bneck，3×3	88	24	-	RE	1
28×28×24	bneck，5×5	96	40	√	HS	2
14×14×40	bneck，5×5	240	40	√	HS	1
14×14×40	bneck，5×5	240	40	√	HS	1
14×14×40	bneck，5×5	120	48	√	HS	1
14×14×48	bneck，5×5	144	48	√	HS	1
14×14×48	bneck，5×5	288	96	√	HS	2
7×7×96	bneck，5×5	576	96	√	HS	1
7×7×96	bneck，5×5	576	96	√	HS	1
7×7×96	conv2d，1×1	-	576	√	HS	1
7×7×576	pool，7×7	-	-	-	-	1
1×1×576	conv2d，1×1，NBN	-	1024	-	HS	1
1×1×1024	conv2d，1×1，NBN	-	k	-	-	1

真实情况	预测结果
真实情况	健康叶片	被感染叶片
健康叶片	TP	FN
被感染叶片	FP	TN