基于动态高效网络的花粉图像识别模型  

Pollen image recognition model based on dynamic and efficient network

作　　者：孙晓川 付多民 秦林林 付景昌 李志刚 SUN Xiao-chuan;FU Duo-min;QIN Lin-lin;FU Jing-chang;LI Zhi-gang(College of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]华北理工大学人工智能学院,河北唐山063210

出　　处：《计算机工程与设计》2023年第3期852-858,共7页Computer Engineering and Design

基　　金：河北省高等学校科学技术研究基金项目(ZD2021088);国家重点研发计划基金项目(2017FE0135700);唐山市科技计划基金项目(19150230E)。

摘　　要：针对花粉图像清晰度不足、形状相近等特性所导致的识别精度低、人工识别低效的问题,提出一种基于动态高效网络的花粉图像识别模型。在ImageNet数据集上使用Noisy Student方法对EfficientNet进行预训练;将训练后的权重迁移到花粉识别模型中;引入动态学习率提升模型的识别精度。采用MixUp和GridMask的方法用于花粉图像数据增强。仿真结果与其它模型进行比较,验证了提出模型具有更好的分类效果。Aiming at the problems of low recognition accuracy and low efficiency of artificial recognition caused by the characteristics of insufficient clarity and similar shape of pollen images,a pollen image recognition model based on dynamic and efficient network was proposed.The noisy student method was used for model pre-training via ImageNet data set.The trained weights were transferred to the pollen recognition model.The dynamic learning rate was applied to improve the recognition accuracy.The Mixup and GridMask methods were used for pollen image enhancement.The simulation results were compared with that of other models.The proposed pollen image recognition model has better classification effects.

关 键 词：花粉图像 深度学习 高效网络 数据增强 嘈杂的学生 迁移学习 动态学习率 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]