基于层间特征蒸馏网络的作物叶片病害检测  

Crop leaf disease detection based on inter‑layer feature distillation network

作　　者：冯玉涵 孙剑 张志芳 Feng Yuhan;Sun Jian;Zhang Zhifang(School of Information Engineering,Xinyang Agriculture and Forestry University,Xinyang,464000,China;College of Computer and Information Technology,Xinyang Normal University,Xinyang,464000,China;College of Electronic Information and Electrical Engineering,Tianshui Normal University,Tianshui,741000,China)

机构地区：[1]信阳农林学院信息工程学院,河南信阳464000 [2]信阳师范大学计算机与信息技术学院,河南信阳464000 [3]天水师范学院电子信息与电气工程学院,甘肃天水741000

出　　处：《中国农机化学报》2024年第9期271-277,共7页Journal of Chinese Agricultural Mechanization

基　　金：河南省科技攻关项目(222102210300,232102210146)。

摘　　要：针对现有农作物叶片病害检测方法对有限标注样本利用不充分,导致模型识别精度不高、泛化性不强的问题,提出一种基于层间特征蒸馏网络的作物叶片病害检测方法。该方法采用支持分支和查询分支相互监督的元学习网络结构,首先,利用一组共享权重的特征提取网络将双分支网络的输入图片映射到深度特征空间,并采用多层下采样操作构造多尺度特征集;然后,在每层特征中计算自注意力机制,在层间计算交叉注意力机制,旨在强化不同尺度内和尺度间特征表达的鲁棒性和可靠性;最后,在跨尺度特征中引入知识蒸馏网络,旨在利用高层特征丰富浅层特征的语义信息,间接地增强不同尺度内和尺度间特征表达的鲁棒性。在马铃薯、苹果、番茄和玉米病害数据集上进行测试,所提方法分别获得0.9531、0.9668、0.9552和0.9542的识别精准率。Aiming at the problem of insufficient utilization of limited labeled samples in existing crop leaf disease detection methods,which leads to low recognition accuracy and weak generalizability of the model,a crop leaf disease detection method based on inter‑layer feature distillation network is proposed.The method adopts a meta‑learning network structure with a support branch and a query branch supervising each other.Firstly,a set of shared weight feature extraction networks are used to map the input images of the two branches to the deep feature space,and multi‑scale feature sets are constructed by using multiple down‑sampling operations.Then,self‑attention mechanism is calculated in each layer feature,and cross‑attention mechanism is calculated between layers,aiming to enhance the robustness and reliability of feature expression at different scales and between scales.Finally,a knowledge distillation network is introduced in the cross‑scale features,aiming to enrich the semantic information of low‑level features with high‑level features indirectly,and further enhance the robustness of feature expression at different scales and between scales.The proposed method has achieved recognition accuracies of 0.9531,0.9668,0.9552 and 0.9542 on potato,apple,tomato and corn diseases,respectively.

关 键 词：病害叶片检测 知识蒸馏 交叉注意力 自注意力 知识反馈 

分 类 号：S435[农业科学—农业昆虫与害虫防治] TP391[农业科学—植物保护]