基于特征学习和关联学习的在线商品跨媒体检索研究

发布时间：2018-01-20 04:01

本文关键词： 跨媒体检索在线商品特征学习关联学习典型相关性分析语义相关匹配　出处：《华东交通大学》2016年硕士论文　论文类型：学位论文

【摘要】：现有电子商务平台仅提供基于关键字匹配的单模态检索,检索准确率偏低,用户常常无法准确地找到所需商品,其检索交互体验较差。跨媒体检索基于多媒体数据之间的语义相关性实现异构媒体之间的直接检索,它能有效提高信息检索的查准率,是单模态检索的有益补充。本文充分考虑商品图像及其文本描述的自身特点,将跨媒体检索模型和方法应用于面向电子商务平台的在线商品检索,主要从特征学习和关联学习两个角度展开研究工作:一、传统文本特征提取方式未考虑关键特征词在文本中出现位置的统计信息,为了突出关键特征词权重,抑制文本噪声干扰,引入Tag-rank模型,根据特征词在文本中的位置统计分布计算其绝对排序和相对排序权重,增强文本特征判别性,最终改善跨媒体检索性能。实验结果表明,在文本检索图像中,绝对排序类模型和相对排序类模型检索性能平均提升4.99%和6.58%;在图像检索文本中,若合理选取图像特征(如GIST特征),也能有效提升检索性能。此外,本文设计的绝对排序类模型和相对排序类模型的输出结果后融合的策略能进一步改善跨媒体检索性能。二、图像和文本之间存在严重的内容鸿沟和语义鸿沟,基于典型相关性分析分别建立核典型相关模型和语义相关匹配模型。核典型相关模型将核函数引入典型相关性分析,学习图像和文本之间隐藏的非线性关联,从而缩小内容鸿沟并且增强两者之间的关联,改善跨媒体检索性能。语义相关匹配模型将典型相关空间的图像和文本特征映射到高层语义空间,并学习两者间的语义关联,以缩小两者之间的语义鸿沟,提高跨媒体检索性能。实验结果表明,基于线性核、高斯核和多项式核的核典型相关模型均能够有效地提升跨媒体检索性能,其中基于高斯核的核典型相关模型平均提升2.37%,效果最显著;语义相关匹配模型能有效地改善文本检索图像性能,平均提高3.82%,在图像检索文本中,若合理地选择图像特征(如SIFT特征),也能改善检索性能。
[Abstract]:The existing e-commerce platform only provides keyword matching based on single mode retrieval, the retrieval accuracy is low, users often can not find the desired goods accurately. Cross-media retrieval based on semantic correlation between multimedia data to achieve direct retrieval between heterogeneous media, it can effectively improve the accuracy of information retrieval. This paper fully considers the characteristics of commodity image and its text description, and applies cross-media retrieval model and method to online commodity retrieval based on e-commerce platform. The main research work is from two aspects of feature learning and association learning: first, the traditional text feature extraction method does not consider the statistical information of the location of key feature words in the text, in order to highlight the weight of key feature words. In order to reduce text noise interference, Tag-rank model is introduced to calculate the absolute ranking and relative sorting weights according to the statistical distribution of the location of feature words in the text, so as to enhance the text feature discrimination. Finally, the performance of cross-media retrieval is improved. The experimental results show that in text retrieval, the retrieval performance of absolute sort model and relative sort class model are improved by 4.99% and 6.58 on average. In the image retrieval text, if the image features (such as GIST features) are properly selected, the retrieval performance can also be effectively improved. The fusion strategy of absolute sort class model and relative sort class model designed in this paper can further improve the performance of cross-media retrieval. Secondly, there is a serious content gap and semantic gap between image and text. Kernel canonical correlation model and semantic correlation matching model are established based on canonical correlation analysis. Kernel canonical correlation model introduces kernel function into canonical correlation analysis to learn the hidden nonlinear correlation between image and text. In order to reduce the content gap and enhance the correlation between the two improve cross-media retrieval performance. Semantic correlation matching model maps the image and text features of the typical correlation space to the high-level semantic space. In order to reduce the semantic gap between the two and improve the performance of cross-media retrieval, the experimental results show that, based on linear kernel. The kernel canonical correlation model of Gao Si kernel and polynomial kernel can effectively improve the performance of cross-media retrieval. Semantic correlation matching model can effectively improve the image performance of text retrieval, with an average increase of 3.82%. In image retrieval, image features (such as SIFT features) should be reasonably selected. It also improves retrieval performance.
【学位授予单位】：华东交通大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】