基于深度学习的服装检索与搭配技术研究

发布时间：2018-03-28 15:09

本文选题：服装检索　切入点：服装搭配　出处：《电子科技大学》2017年硕士论文

【摘要】：随着服装电子商务的发展和大数据时代的来临,互联网上存储着海量的服装数据,用户对服装检索和搭配的需求也日益增加。帮助用户快速精确的找到心仪的服装,并且推荐合适的搭配方案,是服装电商平台提升用户体验和销量的重要手段。正因如此,对服装检索和搭配技术的研究就变得十分有意义。服装检索和搭配技术都依赖于对服装图像内容信息的理解,这种理解就转化为对服装特征的表示上。本文基于深度学习对图像特征的表示,研究了其在服装检索和搭配技术上的应用。本文主要完成以下工作:1.本文介绍了局部特征和深度学习特征两种图像特征的表示方法,分别分析这两种特征的优势和不足,给出了服装检索与搭配技术的理论依据。在局部特征方面,重点介绍了SIFT特征、SURF特征和基于这两种特征编码的BoF特征模型;在深度学习方面,重点介绍了自动编码器和卷积神经网络。2.提出了基于卷积神经网络的服装属性多标签分类模型用于服装图像特征提取。服装图像中含有丰富的服装特有的属性信息,比如颜色、花纹、袖子的长短等等。本文通过训练一个深度卷积神经网络对这些服装属性进行分类,并使用该神经网络的深层激活值表示服装特征。在网络训练过程中,为了弥补服装训练数据的不足,运用迁移学习的方法对网络进行再训练。实验结果表明,该深度卷积网络提取的特征对服装的属性特点能够很好的表示,并且获得了较好的服装检索效果。3.提出了基于相似性度量学习的服装特征优化网络。本文基于Triplet相似性度量学习和双卷积神经网络结构,使用三元组服装图像训练网络参数,对服装特征匹配进行了优化。使用该网络提取服装特征后,用近似最近邻查找相似服装。实验结果表明,优化后的特征在服装商品图和用户拍照图检索效果上都有提升,对服装的光照和形变等干扰因素鲁棒性较高。4.提出了基于服装图像深度特征和局部编码特征的服装搭配空间构建方法。本文基于卷积神经网络提取的特征和SIFT特征编码的BoF特征融合形成服装搭配特征,并使用去噪自动编码器对服装特征降维,然后基于搭配数据集中的搭配频繁项集构建关联规则,使用服装搭配特征和关联规则构建服装搭配空间。实验结果表明,基于该服装搭配空间能够为用户找到合适的服装搭配方案。
[Abstract]:With the development of the clothing e-commerce and big data coming, Internet stores clothing data, users on the increasing demand. Clothing retrieval and collocation to find the right clothing to help users quickly and accurately, and recommend appropriate collocation scheme, clothing is an important means to enhance the user experience and business platform sales. Because of this, the study of clothing collocation retrieval and technology becomes very meaningful. Clothing retrieval and collocation techniques rely on the clothing image content information understanding, this understanding into the clothing features. This paper expressed deep learning representation of image feature based on of clothing and retrieval collocation application technology. This paper mainly completes the following work: 1. this paper introduces the representation of local features and deep learning characteristics of two kinds of image features, analysis of the two respectively. A feature of the advantages and disadvantages of the theory and technology of clothing collocation is given. In the local feature retrieval, introduces SIFT feature, SURF feature and BoF feature model based on the characteristics of two kinds of encoding; learning in depth, focuses on automatic convolution encoder and.2. neural network is proposed for the clothing image feature extraction of clothing attribute convolutional neural network model based on multi label classification. With attribute information, unique rich clothing clothing image such as color, pattern, sleeve length and so on. Through the training of a deep convolutional neural network to classify these clothing attributes, deep and using the neural network's activation value represents the garment feature in the process of network training, in order to compensate for the lack of clothing training data, using the method of transfer learning and training of the network. The experimental results show that the deep The characteristics of attribute extraction convolutional network of clothing can be expressed very well, and get a better retrieval effect.3. presented similar clothing clothing characteristics of network optimization based on metric learning. In this paper, Triplet similarity metric learning and double convolutional neural network structure based on three tuple clothing image training network parameters of garment feature matching was optimized. The extraction of clothing characteristics using the network, using the approximate nearest neighbor searching similar clothing. The experimental results show that the optimized features in clothing product map and user photograph map retrieval effect has improved, the clothing of illumination and deformation of interference factors such as the robust.4. method was proposed to build the depth of clothing image features and local features of clothing collocation based on spatial encoding. This encoding BoF feature and SIFT feature extraction of the convolutional neural network based on Fusion Form the clothing collocation features, and use the denoising auto encoder to reduce the dimensionality of the garment feature, then collocation data set collocation of frequent itemsets constructing based on association rules, the use of clothing collocation features and association rules to construct the clothing collocation space. The experimental results show that the clothing collocation space for users to find suitable clothing collocation scheme based on.

【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】