基于知识蒸馏与EssNet的田间农作物病害识别

Recognizing Crop Diseases in Fields Based on Knowledge Distillation and EssNet

农作物病害的快捷精准识别对我国粮食安全与农业发展提质增效具有重要意义。针对现有病害识别模型参数量大、泛化能力弱、不适用于田间实际场景且不易搭载至移动端等问题,本文提出了EssNet农作物病害识别网络,该网络以ShuffleNetV2_0.5为基础网络,引入高效通道注意力(ECA)机制与SiLU激活函数进行结构改进,同时结合知识蒸馏技术使用EfficientNetB0网络对EssNet进行学习指导,最后使用余弦退火衰减策略对学习率进行动态调整使网络表现达到最优。结果表明,本文提出的EssNet农作物病害识别网络对复杂环境下2种作物(玉米、苹果)的11种病害在测试集上的准确率达到95.21%,比基础网络提高2.11个百分点,参数量为0.35 M,权重文件为1.49 MB。该网络的整体性能优于其他现有模型,为建立田间轻量级农作物病害识别方法提供了参考。

Rapidly and accurately recognizing crop diseases is of great significance for food security and agricultural development in China.Aiming at the problems of the existing disease identification models,such as large parameters,weak generalization ability,not suitable for the actual fields and not easy to carry on the mobile terminal,an EssNet crop disease identification network was proposed in this paper.The network used the ShuffleNetV2_0.5 as basic network,and the efficient channel attention(ECA)mechanism and Sigmoid⁃weighted linear unit(SiLU)activation function were introduced to improve it.At the same time,combined with the theory of knowledge distillation,the EfficientNetB0 was used to guide the learning of EssNet.Finally,cosine annealing was used to dynamically adjust the learning rate to optimize the network performance.The re⁃sults showed that the accuracy of the network on the test dataset of 11 diseases of 2 crops(maize and apple)in fields was 95.21%,which was 2.11 percentage points higher than that of the original network,and the pa⁃rams and weight file size were 0.35 M and 1.49 MB.In conclusion,the proposed EssNet model was better than other existing models,which could provide references for developing lightweight crop disease identification methods.