基于多特征融合的商品图像分类

发布时间：2018-10-15 17:37

【摘要】：由于电子商务网站的成功发展及网络多媒体技术的迅速普及,在线购物已经成为一种方便、快捷、廉价且时尚的购物方式,但随之而来的是图像数据呈几何级数的增长,对如此超大规模的多媒体数据进行有效管理,并提供迅速、准确的检索服务是一个极具挑战性的课题。目前,电子购物网站的搜索服务仍然依赖基于文本的搜索引擎,标注并关联商品的基本信息,对于用户难以准确地描述的样式、花纹、造型等特有属性缺少进一步的标注,将基于内容的图像自动分类引入电子商务,缓解商品图像数据库的管理压力和提高消费者对商品的检索效率,是当前电子商务领域的迫切需求。本文以在线购物商品的图像为基础,构建了一个手工标注商品特殊属性的数据集,并以大量实验关注不同的图像特征对商品图像属性的分类检测结果。主要的研究内容和贡献如下：首先,本文针对原始且粗略的在线商品图像集,从购物用户最关注的色彩和款式两个重要属性出发,基于商品图像特性进行了颜色、纹理和形状分布的剖析,确定运用HSV颜色空间对商品图像提取颜色矩和颜色直方图特征,并采用局部二值模式、梯度局部二值模式、二元梯度轮廓和方向梯度直方图描述纹理信息和形状信息联合表达商品图像的款式属性,通过实验证明了这些特征具有的分类性能。其次,文中详细介绍了不同底层特征对于商品颜色和款式属性的分类方法细节,对两个属性层面的不同特征进行特征级的融合,构建复合的特征向量并通过实验检验特征组合分类的性能变化,实验结果表明,商品图像的分类准确率得到了部分提升。最后,虽然每种特征具备特有的分类性能,但不同特征与分类器决策的相关性没有得到综合利用,采用不同内核的分类算法针对特定特征会有突出的表现,因此本文引入了多内核学习方法改进分类决策,设计和运用大量实验测试了颜色、纹理、形状特征联合描述商品图像属性的能力,对比了多组实验的结果并分析了特征在多核学习中的分类性能。
[Abstract]:With the successful development of e-commerce websites and the rapid popularization of multimedia technology, online shopping has become a convenient, fast, cheap and fashionable way of shopping. It is a challenging task to manage multimedia data on such a large scale effectively and to provide fast and accurate retrieval services. At present, the search service of electronic shopping website still relies on the search engine based on text, marking and associating the basic information of the goods, and lack of further annotation for the unique attributes such as style, pattern, modeling and so on, which are difficult to describe accurately by the user. It is an urgent need to introduce the automatic classification of content-based images into electronic commerce to relieve the management pressure of commodity image database and to improve the retrieval efficiency of consumers in the field of electronic commerce. Based on the images of online shopping items, this paper constructs a data set of manually tagging the special attributes of commodities, and pays close attention to the classification and detection results of commodity image attributes by a large number of experiments. The main research contents and contributions are as follows: first of all, aiming at the original and rough online commodity image set, this paper starts from the two important attributes of color and style that the shopper pays most attention to, and carries on the color based on the commodity image characteristic. Based on the analysis of texture and shape distribution, HSV color space is used to extract color moments and color histogram features from commodity images, and local binary mode and gradient local binary mode are adopted. Binary gradient contour and directional gradient histogram are used to describe texture information and shape information to express the style attributes of commercial images. The classification performance of these features is proved by experiments. Secondly, this paper introduces the classification methods of different bottom features for commodity color and style attributes in detail, and combines the different features of the two attribute levels at the feature level. The experimental results show that the classification accuracy of commodity images has been partially improved. Finally, although each feature has its own classification performance, the correlation between different features and classifier decision is not comprehensively utilized. Therefore, this paper introduces a multi-kernel learning method to improve the classification decision, designs and uses a large number of experiments to test the ability of color, texture and shape features to describe the attributes of commodity images. The results of multi-group experiments are compared and the classification performance of features in multi-core learning is analyzed.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】