摘要
【目的】针对植物病害识别中大规模标注数据依赖性强、新病害适应性差的问题,本研究旨在提升小样本方法在复杂环境下的识别能力,为小样本病害识别提供理论依据。【方法】本研究提出一种结合小样本转换器和监督式掩码知识蒸馏的增强型上下文感知知识蒸馏框架(ECKD),构建包含教师网络与学生网络的协同结构,并结合全局特征对齐、局部特征对齐和监督式掩码图像建模视图策略,引入通道注意力残差模块和上下文感知模块,设计新型ViT-EC编码器以增强模型的特征提取与语义理解能力。基于PlantVillage数据集,采用原型分类器进行5-way 1-shot和5-way 5-shot的小样本任务评估。【结果】在5-way 1-shot和5-way 5-shot小样本任务中,ECKD的原型分类器平均准确率分别达74.98%和88.28%,在多个小样本任务上明显优于现有的主流方法。消融实验表明:注意力残差模块、上下文感知模块和监督式掩码图像建模视图策略对性能有正向贡献,在数据不完整场景下能增强模型对全局语义的理解与局部特征的重建能力。【结论】ECKD通过多视图增强策略与知识蒸馏机制的融合,有效缓解了植物病害中小样本的识别难题,提升了识别的精度和模型的稳定性。该方法为农业智能感知系统提供了高效、可行的解决方案,具有良好的推广潜力。
[Objective]This study aims to address the strong dependence of plant disease recognition on largescale annotated data and its poor adaptability to novel diseases,and enhance the performance of few-shot learning methods under complex environments,so as to provide a theoretical basis for few-shot plant disease recognition.[Method]A novel recognition framework,enhanced context-aware knowledge distillation(ECKD),was proposed by integrating a few-shot transformer with supervised masked knowledge distillation.The framework adopts a collaborative architecture composed of teacher and student networks,and introduces strategies such as global feature alignment,local feature alignment,and supervised masked image modeling views.To enhance feature extraction and semantic understanding,a new encoder ViT-EC was developed using a channel attention residual module and a context-aware module.The model was evaluated using prototype-based classifiers on the PlantVillage dataset under 5-way 1-shot and 5-way 5-shot tasks.[Result]The prototype classifier based on the ECKD framework achieved average accuracies of 74.98%and 88.28%in the 5-way 1-shot and 5-way 5-shot tasks,respectively,significantly outperforming several existing methods.Ablation studies confirmed the positive contributions of the attention residual module,context-aware module,and the supervised masked image modeling strategy in enhancing global semantic understanding and local feature reconstruction,especially under incomplete data conditions.[Conclusion]The ECKD framework effectively alleviates the challenges of few-shot learning in plant disease recognition by integrating multi-view augmentation strategies and a knowledge distillation mechanism.It significantly improves recognition accuracy and model stability,offering an efficient and practical solution for intelligent agricultural perception systems with promising applicability.
作者
聂远
周厚奎
张广群
何涛
胡军国
NIE Yuan;ZHOU Houkui;ZHANG Guangqun;HE Tao;HU Junguo(College of Mathematics and Computer Science,Zhejiang A&F University,Hangzhou 311300,Zhejiang,China;Zhejiang Provincial Key Laboratory of Forestry Intelligent Monitoring and Information Technology,Zhejiang A&F University,Hangzhou 311300,Zhejiang,China;Key Laboratory of Forestry Perception Technology and Intelligent Equipment,National Forestry and Grassland Administration,Zhejiang A&F University,Hangzhou 311300,Zhejiang,China)
出处
《浙江农林大学学报》
北大核心
2025年第4期667-676,共10页
Journal of Zhejiang A&F University
基金
浙江省自然科学基金资助项目(LY24F020005)。
关键词
植物病害识别
小样本学习
知识蒸馏
视图策略
原型分类器
plant disease identification
few-shot learning
knowledge distillation
view strategy
prototype classifier
