您好, 访客   登录/注册

基于深度学习的女衬衫图案样式识别分类

来源:用户上传      作者:李青 冀艳波 郭濠奇 刘凯旋

   摘 要:针对服装图案分类效率低的问题,设计一种基于Inception v3算法与迁移学习技术对女衬衫图案进行分类的方法。在Inception v3基础上,拓展训练网络架构,对8121张8类女衬衫图片进行训练,并与GoogLeNet等典型算法模型进行准确率与损失值对比。结果表明:在相同的识别精度上Inception v3具有较好的收敛速率;并且将迁移学习应用到Inception v3优化算法中,在保持初始模型识别速度情况下,可使模型识别平均精度提高6%,达到98%,同时参与训练的参数量减少了约91%。研究结果可有效解决服装图案分类困难问题,并为服装图案可视化分类研究提供技术参考。
  关键词:Inception v3;迁移学习;衬衫图案;卷积神经网络
  中图分类号:TB383
   文献标志码:A
   文章编号:1009-265X(2022)04-0207-07
  Pattern recognition and classification of women's shirts based on deep learning
  Li Qing1, Ji Yanbo1, Guo Haoqi2, Liu Kaixuan1
  (1.School of Fashion and Art Design, Xi'an Polytechnic University, Xi'an 710048, China;
  2.School of Electrical Engineering and Automation, Jiangxi University of Science andTechnology, Ganzhou 341000, China)
  Abstract: In order to solve the problem of low efficiency of clothing pattern classification, a method based on Inception v3 algorithm and transfer learning technology to classify blouse patterns is designed. On the basis of Inception v3, this paper expand the training network architecture. 8121 pictures of blouses are divided into 8 categories to train, and the accuracy and loss value were compared with typical algorithm models such as GoogLeNet. The results show that: Inception v3 has a better convergence rate at the same recognition accuracy; and transfer learning is applied to the Inception v3 optimization algorithm, while maintaining the initial model recognition speed, the average accuracy of model recognition is increased to 98%, and the amount of parameters involved in training has been reduced by 91%. The results of this research can effectively solve the difficult problem of clothing pattern classification and provide a technical reference for its visual classification research.
  Key words: Inception v3; transfer learning; patterns of blouses ; convolutional neural network
  W购作为新时代下的虚拟平台与实物交易的购物模式,打破了商品区域性,其“一键式”购物逐步被众人接受与应用。服装业呈现出个性化、多元化的发展趋势,并且结合便捷的电商平台,展现出经济蓬勃发展、效益显著提升的局面。在发达的商品化网购时代下,传统特征信息的分类效率与层出不穷的商品供应之间存在不平衡,可表现为商品量大、款式更迭快,但人工分类效率低、主观性较强等方面,导
  致分类速度跟不上新商品涌现的速度。而卷积神经网络CNN[1-2]技术的日益成熟,为服装行业的图像分类、款式生成、图形设计等提供了可行性方案。
  近年来,随着深度学习技术的不断成熟,卷积神经网络在图像分类领域脱颖而出,在服装图像分类方面,张振焕等[3]针对目前服装分类算法在解决多类别服装分类问题时分类精度一般的问题,提出了一种基于残差的优化卷积神经网络服装分类算法;厉智等[4]针对服装图像分类精度较低的问题提出基于深度卷积神经网络的改进服装图像分类检索算法;针对L2-normalization无法跟踪全局信息问题上,Zhe等[5]研究了一种矢量分布的深度度量学习模型,通过定期更迭类中心的替代学习方法,实现了对嵌入层全局信息的捕获和类分布的近似表达;汤清云[6]提出了一种基于注意力区域特征表达的服装图像检索方法;于雨桐[7]利用多特征融合和图像分类识别技术,对服装款式进行明确分类;胡梦莹等[8]提出了利用卷积神经网络对不同品牌服装风格特征进行自动提取、识别和分类的方法实现品牌服装的分类任务;Lü等[9]设计了一种带有姿态预测的深度神经网络模型,提取相关的目标区域特征,并融合分类特征生成服装图像的最终属性。

nlc202209070955



转载注明来源:https://www.xzbu.com/1/view-15439264.htm

相关文章