作物害虫图像智能识别方法  

Image intelligent recognition method of crop pests

作　　者：李雨晴 陈燕红[1] 李永可[1] 肖天赐 李清源 LI Yuqing;CHEN Yanhong;LI Yongke;XIAO Tianci;LI Qingyuan(School of Computer and Information Engineerins,Xinjiang Agricultural University,Urumqi 830052,China)

机构地区：[1]新疆农业大学计算机与信息工程学院,乌鲁木齐830052

出　　处：《新疆农业科学》2023年第12期2973-2981,共9页Xinjiang Agricultural Sciences

基　　金：新疆维吾尔自治区自然科学基金面上项目(2019D01A50);新疆维吾尔自治区重大科技专项课题(2020A01002-4-1)。

摘　　要：【目的】针对作物害虫数据集样本较少、现有单一模型在作物害虫识别上的准确率不高以及泛化能力较差的问题,提出一种基于迁移学习与多模型集成的害虫识别模型。【方法】在大规模公开作物害虫数据集IP102上进行试验,使用迁移学习单独训练6个深层神经网络,选择识别性能较好的EfficientNet、Vision Transformer、Swin Transformer和ConvNeXt进行组合,采用不同策略集成预测结果。【结果】提出的基于迁移学习与多模型集成方法的识别准确率达到75.75%,比性能最好的单模型ConvNeXt提高了1.34%,与目前该数据集上最优算法(CA-EfficientNet)的性能相比,识别准确率高出了6.3%。【结论】害虫图像智能识别模型具有较好的稳定性与泛化能力。【Objective】Aiming at the problems of few crop pest dataset samples,low accuracy of existing single model in crop pest identification and poor generalization ability,a pest identification model based on transfer learning and multi-model integration is proposed.【Methods】Experiments were carried out on the large-scale public crop pest dataset IP102.In this study,transfer learning is used to train 6 deep neural networks separately,and the combination of EfficientNet,Vision Transformer,Swin Transformer and ConvNeXt with better recognition performance was selected,and then different strategies were used to integrate the prediction results.【Results】The results showed that the recognition accuracy of the proposed method based on transfer learning and multi-model integration reached 75.75%,which was 1.34%higher than that of the best-performing single-model ConvNeXt,and was comparable to the current compared with the performance of the optimal algorithm(CA-EfficientNet)on the dataset,the recognition accuracy was 6.3%higher,【Conclusion】It has better stability and generalization ability.

关 键 词：害虫识别 IP102数据集 迁移学习 集成学习 投票法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S41[自动化与计算机技术—计算机科学与技术]