首页 | 本学科首页   官方微博 | 高级检索  
     

基于知识蒸馏与改进ViT网络的花卉图像细粒度分类
引用本文:陈少真,叶武剑,刘怡俊.基于知识蒸馏与改进ViT网络的花卉图像细粒度分类[J].光电子.激光,2024(1):29-40.
作者姓名:陈少真  叶武剑  刘怡俊
作者单位:(广东工业大学 信息工程学院,广东 广州 510006),(广东工业大学 信息工程学院,广东 广州 510006),(广东工业大学 信息工程学院,广东 广州 510006)
基金项目:广东省重点领域研发计划(2018B030338001)、广州市基础研究计划基础与应用基础研究项目 (202201010595)和广东省教育厅创新人才项目和广东工业大学青年百人项目(220413548)资助项目
摘    要:由于自然条件下拍摄的花卉图像背景复杂,而且其存在类内差异性大和类间相似性高的问题,现有主流方法仅依靠卷积模块提取花卉的局部特征难以实现准确的细粒度分类。针对上述问题,本文提出了1种高精度、轻量化的花卉分类方法(ConvTrans-ResMLP),通过结合Transformer模块和残差MLP(multi-layer perceptron) 模块实现对花卉图像的全局特征提取,并在Transformer模块中加入卷积计算使得模型仍保留提取局部特征的能力;同时,为了进一步将花卉分类模型部署到边缘设备中,本研究基于知识蒸馏技术实现对模型的压缩与优化。实验结果表 明,本文所提出的方法在Oxford 17、Oxford 102和自制的Flowers 32数据集上的准确率 分别达98.62%、97.61%和98.40%;知识蒸馏后本文的轻量化模型的大小约为原来的1/18,而准确率仅下降2%左右。因此,本研究能较好地提升边缘设备下花卉细粒度分类的效率,对促进花卉培育的自动化发展具有切实意义。

关 键 词:深度学习    花卉图像分类    自注意力机制    知识蒸馏    迁移学习
收稿时间:2022/7/18 0:00:00
修稿时间:2022/10/21 0:00:00

Flower fine-grained images classification based on the knowledge distillation and improved vision transformer
CHEN Shaozhen,YE Wujian and LIU Yijun.Flower fine-grained images classification based on the knowledge distillation and improved vision transformer[J].Journal of Optoelectronics·laser,2024(1):29-40.
Authors:CHEN Shaozhen  YE Wujian and LIU Yijun
Affiliation:School of Information Engineering, Guangdong University of Technology, Guangzhou, Guangdong 510006, China,School of Information Engineering, Guangdong University of Technology, Guangzhou, Guangdong 510006, China and School of Information Engineering, Guangdong University of Technology, Guangzhou, Guangdong 510006, China
Abstract:Due to the complex background of flower images taken under natural conditions and their high intra-class variability and inter-class similarity,it is difficult to achieve accurate fine-grained classification by existing popular methods relying only on the convolution module to extract local features of flowers.To address the above problems,this paper proposes a high-precision and lightweight flower classification method (ConvTrans-ResMLP).It achieves global feature extraction of flower images by combining the Transformer module and the residual multi-layer perceptron (MLP) module,and adds convolutional computation to the Transformer module so that the model still retains the ability to extract local features.Meanwhile,in order to further deploy the model to edge devices,this study achieves compression and optimization of the model based on knowledge distillation.The experimental results show that the accuracy of proposed method achieves 98.62%,97.61% and 98.40% on Oxford 17,Oxford 102 and homemade Flowers 32 datasets,respectively.The size of the lightweight model in this paper is about 1/18 of the original one after knowledge distillation, while the accuracy rate only decreases by about 2%.Therefore,this study can better improve the efficiency of flower fine-grained classification by edge equipment, which is of practical significance to promote the automation of flower cultivation.
Keywords:deep learning  flower image classification  self-attention mechanism  knowledge distillation  transfer learning
点击此处可从《光电子.激光》浏览原始摘要信息
点击此处可从《光电子.激光》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号