【作者】 高珊；

【导师】 陈纯； 宋明黎；

【作者基本信息】 浙江大学 ， 计算机应用技术， 2016， 硕士

【摘要】 随着电子商务的发展,如何正确识别服装风格成为一项意义深远的工作。服装风格的准确识别可为如下研究提供基础：自动化的服装标注、基于内容的服装检索、个性化的服装推荐、风格统一的服装搭配。而现有服装风格识别方法主要基于客观风格,较少基于主观风格,且缺乏对多个主观风格同时识别的工作,因此本文提出一种基于多个主观风格的服装识别方法。我们首先从互联网上爬取多张带有多个主观风格的服装图片及其主、客观风格信息；然后对这些图片进行姿势估计和特征提取；其次为了能够得到更有区分度的特征,我们提出一种多标签距离度量学习模型,并将该模型运用到已获得的特征上,得到新的距离度量学习特征。之后为了进一步提高服装主观风格的识别效果,我们结合服装图片的客观风格文本描述信息,将已有的单标签多视图学习模型改进成为多标签多视图学习模型,得到多视图学习的新特征。最后对多视图学习特征集进行多标签分类,得到服装多个主观风格的识别结果。实验结果表明,本文提出的基于距离度量学习和多视图学习的服装主观风格识别方法可以有效地识别服装主观风格。更多还原

【Abstract】 With the development of electronic commerce, more and more people are interested in buying clothes online. In the field of computer vision, how to recognize the styles of clothes becomes a piece of meaningful work. The accurate identification of clothing styles can lay a good foundation for the following fields:automatic labelling, retrieval based on the content of the clothing styles and individual recommendation, and even collocating. However, existing methods mostly focus on objective styles, rather than subjective ones. And the problem of identifying multiple subjective styles has rarely been addressed before. Therefore, this paper puts forward a method to identify multiple clothing subjective styles at the same time.In this paper, we crawl many clothing images with manual labeled multiple subjective styles and their whole label information from the Internet; Then do pose estimation and extract features on these images; After that, in order to get more suitable features, we propose a multi-label distance metric learning model inspired by single-label distance metric learning. Next, in order to increase the effectiveness of clothing subjective styles recognition using the text information, we put forwards a multi-label multi-view learning model, and this model is applied to features obtained by last step. Finally, we adopt multi-label classification method on the features learned from the above models, and improve the performances in subject clothes styles recognition.The experimental results show that the proposed method can effectively recognize the subjective styles of the clothing images.更多还原