基于热点标签学的可视化时尚商品搜索引擎设计与实现

发布时间：2018-09-06 09:35

【摘要】：本论文研究课题来源于广州电信研究院“基于云计算的大数据应用研究”子课题之一，主要研究大数据处理(Big Data)技术如分布式存储、分布式计算在服装频道上针对数据处理领域的应用。随着互联网日益普及、电子商务的发展，在线购物已渗透到人们的日常生活。目前主流购物平台主要向用户提供文本检索模式，但由于某些商品复杂的外观以及文字描述的不确切性，导致搜索结果往往不令人满意，如外观复杂的服饰。由此可见，“以图搜图”已成为一项迫切的检索模式。尽管也有少量平台尝试提供依靠图片检索商品，但是受图片背景干扰严重，同样不能达到较好的检索效果。针对上述问题，本论文针对服装频道提出了一种基于热点标签自动学习的检索模型，属于“以图搜图”范畴。该模型基于互联网时尚商品仅需少量热点文本标签即可得到相对令人满意的搜索结果这样一个事实设计并实现，具体而言，即通过识别用户提交图片中包含的时尚元素热点标签，检索出具有相同热点标签的图像集，完成一个从图到热点标签，再从热点标签到图的检索过程。包括训练阶段和搜索阶段。在训练阶段，从互联网收集一定量半标注的包含热点标签的图片形成训练集，提取图像的时尚元素特征，结合SVM生成热点标签分类器；在搜索阶段，对提交查询的图片进行热点标签提取，然后利用热点标签从商品库中搜索并返回结果。本论文提出模型的优点主要包括两点：一是相比现有可视化搜索引擎，该方法具有不易受背景图片干扰，对时尚商品针对性更强；二是具有一定的语义搜索效果。本论文的主要创新点在于：提出了一种从图像到热点标签再到图像的搜索模型，发挥热点标签学习相对容易并且具有很强表达能力的优势。本论文主要贡献有四点：一是提供了可用于进一步学习研究的半标注图像数据库，，包括三种领子样式和三种纹理样式。二是结合Nutch开源项目，完成了具有通用性的网络数据抓取模块。三是结合Hadoop分布式技术，收集了一个可用于图像检索技术研究的图像数据集。四是按照软件工程的要求实现了支持本论文方法的原型系统，用户仅需提交一张图片即可搜索到感兴趣的结果。
[Abstract]:This dissertation is one of the sub-topics of "Application Research of Large Data Based on Cloud Computing" of Guangzhou Telecommunication Research Institute. It mainly studies the application of Big Data technology such as distributed storage and distributed computing in the field of data processing on clothing channel. Shopping has penetrated into people's daily life. At present, the mainstream shopping platforms mainly provide users with text retrieval mode, but because of the complex appearance of some goods and the inaccuracy of text description, the search results are often unsatisfactory, such as clothing with complex appearance. Despite a small number of platforms attempting to provide commodities that rely on image retrieval, but seriously disturbed by the background of the image, the same can not achieve better retrieval results.
Aiming at these problems, this paper proposes a retrieval model based on hot tag automatic learning for fashion channel, which belongs to the category of "search by graph". This model is based on the fact that fashion products on the Internet can get relatively satisfactory search results with only a few hot text tags. Specifically, it is By identifying the hotspot tags of fashion elements contained in the pictures submitted by users, the image set with the same hotspot tags is retrieved, and a retrieval process is completed from the graph to the hotspot tag, and then from the hotspot tag to the graph, including the training stage and the search stage. Pictures form a training set, extract the fashion element features of images, and generate a hot label classifier with SVM. In the search phase, hot labels are extracted from the submitted images, and then hot labels are used to search the merchandise stores and return the results.
The advantages of this model include two main points: first, compared with the existing visual search engine, this method is not easily disturbed by background pictures, more targeted to fashion products; second, it has a certain semantic search effect. The main contributions of this paper are as follows: Firstly, we provide a semi-annotated image database which can be used for further study, including three collar styles and three texture styles; secondly, we complete the general purpose with Nutch open source project. The third is to collect an image data set which can be used in the research of image retrieval technology with the help of Hadoop distributed technology. The fourth is to implement a prototype system which supports the method of this paper according to the requirement of software engineering.
【学位授予单位】：中山大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【相似文献】