面向购物搜索的目标提取算法研究及系统实现

发布时间：2018-03-18 21:26

本文选题：购物图像搜索　切入点：商品提取　出处：《西南交通大学》2012年硕士论文　论文类型：学位论文

【摘要】：由于电子商务网站的成功发展,在线购物已经成为一种方便、快捷、廉价的购物方式,随之而来的是图像数据呈现几何级数增长,如何对如此超大规模的购物图像进行有效搜索成为近年来学术界和工业界的研究热点。目前Google和阿里巴巴等已经提供了查找视觉相似商品的搜索服务,但是这些搜索引擎由于对复杂背景的图像直接提取形状、纹理、颜色等的全局视觉特征而受到图像背景噪声的干扰,因此无法取得理想的搜索效果。为了提高购物图像的搜索准确度,必须去除图像的复杂背景,即提取出图像中的商品目标。本文针对包和衣服类复杂背景购物图像的商品提取问题,提出了与传统机器学习不同的检查方法,旨在用以提高购物图像的搜索准确度。本文的主要内容和贡献如下：第一,提出了购物图像主目标提取算法,该算法主要针对不含有模特的图像提取商品目标。购物图像虽然背景复杂却有这样的特点,商品目标一般被置于接近图像中心的位置,并且目标对象应该占图像足够比例以醒目。于是通过利用基于图的快速分割算法对图像对象的识别能力,以及主目标的空间位置分布特性和区域大小特性,本文提出了与传统机器学习不同的检测方法算法来获取目标对象。第二,提出了购物图像多目标提取算法,该算法主要处理含有模特的购物图像。购物图像中的模特一方面为商品提取增加了难度,另一方面也为找到衣物提供了线索。该算法首先利用人脸和肤色等的先验知识大致定位可能的衣物区域；接着根据高斯混合模型分析了图像的背景和衣物模型,并加入空间信息修正这些模型；最后根据模型准确地得到衣物。第三,实现了购物搜索系统。利用主目标、多目标提取算法去除购物图像的背景干扰后再提取图像的颜色、SIFT (Scale Invariant Feature Transform)特征,前台采用Grub Cut分割算法与用户交互,最后利用欧式距离和BoW (Bag of Words)分别匹配颜色和SIFT特征。通过实验一方面证明了两种提取算法的有效性,另一方面说明本文的搜索系统能提高购物图像的检索准确度。
[Abstract]:Due to the successful development of e-commerce websites, online shopping has become a convenient, fast and cheap way of shopping, followed by the geometric growth of image data. How to effectively search such large scale shopping images has become a hot research topic in academia and industry in recent years. At present, Google and Alibaba have provided search services to find visual similar products. However, these search engines are disturbed by background noise because of extracting the global visual features of shape, texture, color and so on directly from the image of complex background. In order to improve the search accuracy of the shopping image, the complex background of the image must be removed. In this paper, a different checking method from traditional machine learning is proposed to extract commodities from shopping images with complicated background of bags and clothes. This paper aims to improve the search accuracy of shopping images. The main contents and contributions of this paper are as follows:. First, the main object extraction algorithm of shopping image is proposed. The algorithm is mainly aimed at extracting commodity target without models. Although the background of shopping image is complex, it has such characteristics. Commodity objects are generally placed close to the center of the image, and the target object should account for a sufficient proportion of the image to stand out. Thus, the ability to recognize the image object is achieved by using a fast segmentation algorithm based on a graph. As well as the spatial distribution characteristics and region size characteristics of the main target, this paper proposes a different detection algorithm from the traditional machine learning algorithm to obtain the target object. Secondly, a multi-object extraction algorithm for shopping images is proposed, which mainly deals with shopping images with models. On the one hand, models in shopping images make it more difficult to extract goods. On the other hand, it also provides clues to find clothing. Firstly, the algorithm uses the prior knowledge of face and skin color to roughly locate the possible clothing region. Then, according to Gao Si mixed model, the background and clothing model of the image are analyzed. These models are modified by adding spatial information. Finally, the clothing is accurately obtained according to the model. Thirdly, a shopping search system is implemented, which uses the main target, multi-target extraction algorithm to remove the background interference of the shopping image, and then extracts the color sift scale Invariant Feature transform feature of the image. The foreground uses Grub Cut segmentation algorithm to interact with the user. Finally, the Euclidean distance and BoW bag of Wordsare used to match the color and SIFT features respectively. On the one hand, the validity of the two extraction algorithms is proved by experiments, on the other hand, the search system in this paper can improve the retrieval accuracy of shopping images.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【参考文献】