基于EfficientNet改进模型的服饰图像智能分类技术  

Intelligent classification technology for clothing images based on improved EfficientNet model

作　　者：王佳鑫 李雪飞 张颖 WANG Jiaxin;LI Xuefei;ZHANG Ying(Information Center,Beijing Institute of Fashion Technology,Beijing,China;School of Fashion Accessory,Beijing Institute of Fashion Technology,Beijing,China)

机构地区：[1]北京服装学院信息中心,北京 [2]北京服装学院服装艺术与工程学院,北京

出　　处：《东华大学学报（自然科学版）》2024年第6期151-157,共7页Journal of Donghua University(Natural Science)

基　　金：2020北京市教委项目(KM202010012007);北京服装学院研究生科研创新项目(X2024-124)。

摘　　要：为了提高服饰图像的智能分类效率和分类准确性,提出一种基于EfficientNet模型改进的服饰分类模型。该模型基于卷积神经网络构建,将EfficientNet-B0模型和EfficientNet-B1模型与SE模块相结合,对卷积神经网络的结构和激活函数进行改进,以提升其特征提取和表达能力。结果显示,改进后的模型不会引入大量参数,且改进后整体上模型在服饰数据集UT-zappos50K和Fashion-MNIST数据集上的分类准确率相较于VGG16、Swin Transformer等经典卷积神经网络模型最多提高了2.65%。这表明该改进方式能够有效提高服饰图像分类模型的性能。To enhance the intelligent classification efficiency and accuracy of clothing images,a clothing classification model is proposed,based on the EfficientNet model.Constructed using convolutional neural networks,this model integrates and combines the EfficientNet-B0 and EfficientNet-B1 models with SE modules.It also improves the structure and activation functions of the convolutional neural network to enhance its feature extraction and representation capabilities.The results demonstrate that the improved model,without introducing a significant number of parameters,achieves a maximum increase of 2.65%in classification accuracy on clothing datasets such as UT-zappos50K and Fashion-MNIST,as compared to classic convolutional neural network models like VGG16 and Swin Transformer.This validates that this improvement approach is effective in enhancing the performance of clothing image classification models.

关 键 词：深度学习 神经网络 图像分类 激活函数 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]