基于改进Res2Net与迁移学习的水果图像分类  

Fruit image classification based on improved Res2Net and transfer learning

作　　者：吴迪[1,2] 肖衍[1] 沈学军 万琴[1,2] 陈子涵 WU Di;XIAO Yan;SHEN Xuejun;WAN Qin;CHEN Zihan(Institute of Electrical and Information Engineering,Hunan Institute of Engineering,Xiangtan 411100,China;National Engineering Research Center of RVC,Hunan University,Changsha 410082,China)

机构地区：[1]湖南工程学院电气与信息工程学院,湘潭411100 [2]湖南大学机器人视觉感知与控制技术国家工程研究中心,长沙410082

出　　处：《电子科技大学学报》2025年第1期62-71,共10页Journal of University of Electronic Science and Technology of China

基　　金：国家重点研发计划(2020YFB1713600);国家自然科学基金(62476084);湖南省教育厅重点项目(24A0528);湖南省自然科学基金(2022JJ30198);湖南省研究生科研创新项目(YC202228)。

摘　　要：针对传统水果图像分类算法特征学习能力弱和细粒度特征信息表示不强的缺点,提出一种基于改进Res2Net与迁移学习的水果图像分类算法。首先,针对网络结构,在Res2Net的残差单元中引入动态多尺度融合注意力模块,对各种尺寸的图像动态地生成卷积核,利用meta-ACON激活函数优化ReLU激活函数,动态学习激活函数的线性和非线性,自适应选择是否激活神经元;其次,采用基于模型迁移的训练方式进一步提升分类的效率与鲁棒性。实验结果表明,该算法在Fruit-Dataset和Fruits-360数据集上的测试准确率相比Res2Net提升了1.2%和1.0%,召回率相比Res2Net提升了1.13%和0.89%,有效提升了水果图像分类性能。Aiming at the shortcomings of the traditional fruit image classification algorithm with weak feature learning ability and weak representation of fine-grained feature information,this paper proposes a fruit image classification algorithm based on improved Res2Net with migration learning.First,for the network structure,a dynamic multi-scale fusion attention module is introduced into the residual unit of Res2Net to dynamically generate convolution kernels for images of various sizes,optimize the ReLU activation function by using the meta-ACON activation function,and dynamically learn the linearity and nonlinearity of the activation function to adaptively choose whether to activate the neurons or not;second,a training method based on model migration is used to further improve the efficiency and robustness of classification.The experimental results show that the algorithm proposed in this paper improves the test accuracy on Fruit-Dataset and Fruits-360 dataset by 1.2%and 1%compared with Res2Net,and the recall rate improves by 1.13%and 0.89%compared with Res2Net,which effectively improves the performance of fruit image classification.

关 键 词：图像分类 Res2Net 动态多尺度融合注意力 激活函数 迁移学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]