面向图像购物搜索的图像分析方法研究

发布时间：2018-04-04 22:05

本文选题：SIFT描述子　切入点：词袋模型BoW　出处：《南京财经大学》2016年硕士论文

【摘要】：电子商务的迅猛发展改变了人们传统的购物习惯。然而,现有的利用分类搜索和关键字搜索的商品搜索技术,存在搜索结果量大而匹配精确度低的问题。本文以商品图像搜索为研究焦点,重点研究了其核心问题-商品图像的特征抽取和匹配问题。本文工作的主要贡献是针对经典的尺度不变特征变换描述子SIFT(Scale Invariant Feature Transform),在应用于商品图像识别中,存在的在较大仿射变换和视角变换的情况下,无法进行有效匹配的问题,提出了一种新的融合多视角的仿射不变描述子。该描述子首先将商品图像进行模拟视角转换,生成一组模拟视角图像序列,然后检测序列图像的视觉特征,最后利用随机抽样一致算法RANSAC(Random Sample Consensus),将模拟视角图像序列中的视觉特征映射到原始图像中,共同构成原始图像的特征点。本文还对传统词袋模型BoW(Bag-of-Words)中视觉词典的构造方法和视觉特征的量化方法进行了改进。针对构造视觉词典的传统K-Means聚类算法存在的初始聚类中心的随机选取,所导致的聚类结果不稳定,且易存在局部极值点的问题,提出采用密度敏感相似性度量方法确定K-Means聚类算法的初始聚类中心;并针对传统视觉特征量化采用硬量化方法HQ(Hard Quantization),而忽视了视觉特征与视觉单词之间联系的问题,提出采用软量化方法SQ(Soft Quantization)方法进行视觉特征量化。用从淘淘搜、最美搜衣、唯品会等购物网站上抓取的13000幅服装商品的图像构造商品图像数据库,对本文提出的方法进行测试,并与采用梯度方向直方图HOG(Histogram of Oriented Gradient),常规SIFT特征以及传统BoW的方法进行比较,证明了本文提出的图像特征提取算法和改进的BoW算法的有效性。
[Abstract]:The rapid development of electronic commerce has changed people's traditional shopping habits.However, the existing commodity search techniques using classified search and keyword search have the problem of large amount of search results and low matching accuracy.In this paper, we focus on commodity image search, and focus on the feature extraction and matching of commodity image.The main contribution of this paper is that the classical scale-invariant feature transform descriptor SIFT(Scale Invariant Feature transform can not match effectively when it is applied to commodity image recognition under the condition of large affine transformation and angle of view transformation.A new affine invariant descriptor is proposed.The descriptor first transforms the commodity image into a set of simulated visual angle images, and then detects the visual features of the sequence images.Finally, a random sampling algorithm, RANSAC(Random Sample Consensusn, is used to map the visual features of the simulated visual angle image sequence to the original image, so as to form the feature points of the original image.This paper also improves the method of constructing visual dictionaries and quantifying visual features in the traditional word bag model BoW-Bag-of-Words'.According to the random selection of the initial clustering center in the traditional K-Means clustering algorithm, the clustering results are unstable and the problem of local extremum is easy to exist.The initial clustering center of K-Means clustering algorithm is determined by using density-sensitive similarity measure, and HQ(Hard quantization method is used for traditional visual feature quantization, which neglects the connection between visual features and visual words.A soft quantization method (SQ(Soft quantization) is proposed for visual feature quantization.Using the images of 13000 clothing items seized on shopping websites such as Amoy search, most Beautiful clothing search, VIPSHOP and so on, to construct a commodity image database, and to test the methods proposed in this paper.Compared with the methods of HOG(Histogram of Oriented gradient histogram, conventional SIFT features and traditional BoW, the effectiveness of the proposed image feature extraction algorithm and the improved BoW algorithm is proved.
【学位授予单位】：南京财经大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】