复杂田间环境下基于ResNet18-GABM的马铃薯叶片主要病害自动识别方法

A ResNet18-GABM-based method for identifying major potato leaf diseases in complex field environments

关键词:  马铃薯病害识别  复杂背景叶片  迁移学习  注意力机制  病害严重度

Key words:  potato disease recognition  leaves with complex backgrounds  transfer learning  attention mechanism  disease severity

摘要:

【目的】 马铃薯早疫病和晚疫病频发，严重情况下可导致产量下降10%~40%，对粮食安全构成较大威胁。文章针对现有病害识别模型在复杂田间环境下准确率不足的问题，构建了一种面向真实场景的马铃薯病害分类与分级模型ResNet18-GABM。【方法】 该模型以ResNet18为骨干网络，引入全局分组坐标注意力模块（GGCA），增强对病斑细微特征的表征能力。同时设计多尺度特征双时融合模块（BFM），实现多维语义信息的有效融合，并结合迁移学习策略优化训练过程，在提升收敛效率的同时提高识别精度和计算效率。基于单叶片、多叶片和混合叶片3类复杂场景数据集开展消融与对比实验，并对不同严重程度的病害进行分级识别。【结果】 （1）将GGCA与BFM模块联合嵌入ResNet18模型后，其分类准确率较基线模型提升2.09个百分点。（2）该模型在单叶片数据集上的识别准确率为100%；在多叶片数据集上的准确率为95.08%，较基准模型ResNet18提高4.92个百分点；在混合叶片数据集上的分类精度为99.10%，较4类基线模型提升2.09~3.59个百分点。（3）在严重程度分级任务中，该模型的识别准确率达到92.66%，较基线ResNet18提升2.75个百分点。【结论】 该模型在复杂农田环境下具有良好的泛化能力和稳定性，可为马铃薯病害的早期预警与精准防控提供技术支持。

Abstract:

[Purpose] Potato early and late blights occur frequently，potentially leading to yield losses of 10% to 40% in severe cases，which poses a significant threat to food security. To address the insufficient accuracy of existing disease recognition models in complex field environments，this study constructs a potato disease classification and grading model，ResNet18-GABM，which is designed for real-world scenarios.[Method] Using ResNet18 as the backbone network，the model integrated a global grouped coordinate attention （GGCA）module to enhance the representation capability of subtle lesion features. Meanwhile，a multi-scale Bi-temporal fusion module （BFM）was designed to achieve effective fusion of multi-dimensional semantic information，and a transfer learning strategy was employed to optimize the training process，improving convergence efficiency while enhancing recognition accuracy and computational performance. Ablation and comparative experiments were conducted using three types of complex scene dataset：single-leaf，multi-leaf，and mixed-leaf. Grading recognition was then performed for diseases of varying severity levels.[Result] （1）The integration of GGCA and BFM modules into ResNet18 yielded a 2.09 percentage point increase in classification accuracy over the baseline. （2） The proposed model achieved a perfect recognition rate of 100% on single-leaf samples. On the multi-leaf dataset，its accuray was 95.08%，outperforming the standard ResNet18 by 4.92 percentage points. Furthermore，the classification precision on the mixed-leaf dataset reached 99.10%，representing a gain of 2.09%-3.59% compared to four state-of-the-art baseline models. （3）For the disease severity grading task，the model attained an accuracy of 92.66%，representing a 2.75 percentage point improvement over the original ResNet18.[Conclusion] The proposed model demonstrates robust generalization capabilities and stability in complex agricultural environments，providing technical support for the early warning and precise prevention and control of potato diseases.