DB-DAB-Net：一种用于玉米叶片病害识别的高效新型网络

DB-DAB-Net: An Efficient Novel Network for Maize Leaf Disease Identification

摘要

玉米是全球重要的粮食作物，其生产过程易受各种病害的侵袭，在复杂背景下，由于现有的深度学习模型具有多尺度特征提取能力不足和上下文信息融合效率低等缺陷，易导致病害识别准确率低。为此，本文提出一种双分支特征提取结构与注意力机制融合的玉米叶片病害识别模型（DB-DAB-Net），首先构建深度可分离卷积（Depthwise Separable Conv）的双分支特征提取结构，提取复杂特征和空间细节，并降低计算量和参数量，利用动态融合门控机制自适应双分支，提升特征提取效率和模型鲁棒性。其次，在每个阶段后加入融合通道注意力和空间注意力的MCSA（Multi-scale Combination-Selective Attention）模块，并引入双向交互权重机制，实现对全局和局部信息的精细化捕捉，降低计算成本。最后，在平均池化层前加入BiFPN（Bidirectional Feature Pyramid Network，双向特征金字塔网络）模块，通过双向特征金字塔网络进行不同尺度的特征融合，使用可学习权重动态调整各尺度特征，并采用Swish激活函数保证梯度稳定性，进一步提升模型对多尺度目标的感知能力。实验结果表明，DB-DAB-Net在玉米病害数据集上具有良好的识别性能，识别查准率、召回率、F1分数、准确率分别达到了97.58%、97.47%、97.49%、97.47%，同时模型的参数量和浮点计算量分别为2.53 M和5.56 G。在复杂环境下，DB-DAB-Net模型能够有效提高玉米叶片病害识别的准确性，同时为农业病害监测提供了一种新的技术思路。

Abstract

Maize is a vital global food crop, yet its production is vulnerable to various diseases. Under complex backgrounds, the existing deep learning models often suffer from insufficient multi-scale feature extraction and inefficient fusion of contextual information, leading to low accuracy in disease identification. To address these issues, this paper proposes a maize leaf disease recognition model (DB-DAB-Net), which combines a two-branch feature extraction structure with the attention mechanism. Firstly, Depthwise Separable Conv is constructed as a two-branch feature extraction structure to capture complex features and spatial details while reducing computational cost and parameter count. A dynamic fusion gating mechanism is introduced to adaptively integrate the double branch, enhancing the feature extraction efficiency and model robustness. Then, MCSA module is incorporated after each stage, which combines channel attention and spatial attention. A bidirectional interactive weight mechanism is introduced to achieve fine capture of global and local information while minimizing calculation cost. Finally, BiFPN module is added in front of the average pooling layer, which performs multi-scale feature fusion through the bidirectional feature pyramid network, dynamically adjusts features at different scales using learnable weights, and employs Swish activation function to ensure gradient stability, so as to further improve the model's perception ability of multi-scale targets. The experimental results show that DB-DAB-Net achieves good identification performance on maize disease data set, with the identification accuracy rate, recall rate, F1 score and accuracy rate reaching 97.58%、97.47%、97.49% and 97.47%, respectively. The number of parameters and floating-point operations of the model are 2.53 M and 5.56 G, respectively. In complex environment, DB-DAB-Net model can effectively improve the accuracy of maize leaf disease detection, providing a new technical idea for agricultural disease monitoring.

关键词

玉米叶片病害 / 双分支网络 / 多尺度特征融合 / 注意力机制 / 智慧农业

Key words

Maize leaf disease / double-branch network / multi-scale feature fusion / attention mechanism / smart agriculture

参考文献

[1] 张辉. 施用污泥和秸秆堆肥对盐碱土壤理化性质的影响研究[J]. 大众科技, 2025, 27(01): 68-70+91.

[2] 单长林, 李孝军, 周圆, 等. 进境玉米穗腐病菌的分离鉴定[J]. 植物病理学报, 2024, 54(04): 852-856.

[3] 熊诗洁. 花椒精油及其主要成分对番茄果腐病的抑制及机理[D].荆州:长江大学, 2024.

[4] 蔡广兴, 王丽娟, 刘可杰, 等. 玉米农家品种对玉蜀黍尾孢菌引起的灰斑病抗性鉴定[J]. 辽宁农业科学, 2025(02): 69-71.

[5] Mehta S , Kukreja V , Gupta A . Revolutionizing maize disease management with federated learning CNNs: A decentralized and privacy-sensitive approach[C] 2023 4th International Conference for Emerging Technology (INCET). IEEE, 2023: 1-6.

[6] 马春悦, 郭秀茹, 王琛, 等. 一种基于改进ResNet18神经网络的玉米叶片病害识别方法[J]. 山东农业大学学报(自然科学版), 2024, 55(03): 356-366.

[7] 曾鹏滔, 撒金海, 刘嘉. 基于改进ResNet18 的玉米叶片病害识别方法[J]. 内蒙古农业大学学报(自然科学版), 2023, 44(06): 60-67.

[8] Ramadan S T Y , Sakib T , Jahangir R , et al. Maize leaf disease detection using vision transformers(ViTs) and CNN-Based classifiers:Comparative analysis[C]. International Conference on Human-Centric Smart Computing. Singapore: Springer Nature Singapore, 2023: 513-524.

[9] Amrani A , Sohel F , Diepeveen D , et al. Insect detection from imagery using YOLOv3-based adaptive feature fusion convolution network[J]. Crop and Pasture Science, 2022, 6: 615-627.

[10] 李新科, 毛含莹, 孙玉瑾, 等. 基于DCNN的水稻叶瘟病检测与取样装置设计实现[J]. 电子制作, 2025, 33(06): 36-41.

[11] 党宏社, 肖利霞, 张选德. 不良光照场景下的交通标志识别算法[J]. 半导体光电, 2025, 46(01): 142-148.

[12] 戴佩玉, 张欣, 毛星, 等. 利用空间-光谱双分支特征和动态选择的高光谱影像农作物分类[J]. 农业工程学报, 2023, 39(16): 160-170.

[13] 骆津津, 陈伟, 田子建, 等. 基于YOLOv8-ECW的井下人员行为实时检测算法[J]. 矿业科学学报, 2025, 10(02): 316-327.

[14] Woo S, Park J, Lee JY, et al. Cbam: Convolutional block attention module[C]. Proceedings of the European Conference on Computer Vision(ECCV), 2018: 3-19.

[15] 徐沁, 梁玉莲, 王冬越, 等. 基于 SE-Res2Net 与多尺度空谱融合注意力机制的高光谱图像分类[J]. 计算机辅助设计与图形学学报, 2021, 33(11): 1726-1734.

[16] Ni H, Shi Z, Karungaru S, et al. Classification of typical pests and diseases of rice based on the ECA attention mechanism[J]. Agriculture, 2023, 13(5): 1066.

[17] Ouyang D , He S , Zhang G , et al. Efficient multi-scale attention module with cross-spatial learning[C]. ICASSP 2023-2023 IEEE International Conference on Acoustics,Speech and Signal Processing (ICASSP). IEEE, 2023: 1-5.

[18] Chen J , Mai HS , Luo L , et al. Effective feature fusion network in BIFPN for small object detection[C]. 2021 IEEE international conference on image processing (ICIP).IEEE, 2021: 699-703.

[19] Qiu M, Huang L, Tang BH . ASFF-YOLOv5: Multielement detection method for road traffic in UAV images based on multiscale feature fusion[J]. Remote Sensing, 2022, 14(14): 3498.

[20] Nawaz SA, Li J, Li D, et al. Medical image zero watermarking algorithm based on dual-tree complex wavelet transform, AlexNet and discrete cosine transform[J]. Applied Soft Computing, 2025, 169: 112556.

[21] Subba AB , Sunaniya AK . Computationally optimized brain tumor classification using attention based GoogLeNet-style CNN[J]. Expert Systems with Applications, 2025, 260: 125443.

[22] Tanwar N, Turukmane AV . Modified MobileNetV2 transfer learning model to detect road potholes[J]. PeerJ Computer Science, 2025, 11: e2519.

[23] Liu H, Dou Y, Wang K, et al. A skin disease classification model based on multi scale combined efficient channel attention module[J]. Scientific Reports, 2025, 15(1): 6116.

[24] Dai G ,Fan ,JTian Z, et al. PPLC-Net:Neural network-based plant disease identification model supported by weather data augmentation and multi-level attention mechanism[J]. Journal of King Saud University-Computer and Information Sciences, 2023, 35(5): 101555.

[25] 代国威. 基于多层次深度学习的马铃薯叶片病害诊断模型构建与优化[D].北京:中国农业科学院, 2024.

[26] 李广睿, 刘琼, 张熠卿, 等. 面向安检X 光图像的违禁物品语义分割与识别研究[J]. 电子测量与仪器学报, 2024, 38(02): 1-9.

基金资助

国家自然科学基金(31870532)

长沙市科技计划项目(kq2402265)