基于深度学习的视觉特征在图像检索中的应用-计算机应用技术专业论文.docx

The deep learning of visual features in image search applications A Dissertation Submitted to the Graduate School of Henan University in Partial Fulfillment of the Requirements for the Degree of Master of Engineering Science By Guo Xin Mentor：Prof. Wang GuoAn May,2015 I I 摘 要 当前的图像检索已经由单一的基于文本或内容检索逐步向基于文本、内容、用户倾 向以及视觉需求等综合特征的方向发展。并且随着大数据时代到来，基于统计原理的各 种机器学习算法得以广泛应用。尤其是基于深度学习的视觉特征以及 LTR 排序模型的 使用，使图像检索技术得到空前的发展。 本文介绍基于深度学习的视觉特征的一个重要应用，对视觉相似度特征的基本概念 以及相关生成算法做了详细描述，包括其中视觉特征的稀疏矩阵优化方案。然后对视觉 相似度特征生成的整个架构支持的基础进行介绍。并详细描述了架构实现的技术点，以 及策略算法实现。最后展示了该视觉相似度特征在图像检索系统中的实际应用以及带来 的检索效果的提升。 之后简要介绍基于深度学习的视觉特征的另外两个应用。视觉相似反馈 Outlier 策 略以及基于 CDNN 网络的视觉文本特征深度学习。并分小节介绍了两个应用的具体算 法步骤以及在实际检索效果中增加收益。这些应用都是基于深度学习的视觉特征的延 伸，其中基于 CDNN 网络的视觉文本特征深度学习，更是对视觉特征产生的方法进行 了扩展，将深度学习这种最先进的机器学习思想应用到了 LTR 排序模型特征的生成上， 并取得了成功。 关键词：图像检索，机器学习，计算机视觉，深度学习，GBrank III III ABSTRACT The current image retrieval has developed from a single step by step to the development of comprehensive character of the text, content, user orientation and visual demand based on the direction of text or content based retrieval. And as the big data era, a variety of statistical machine learning algorithms can be widely applied principle based on. Especially the use of visual features deep learning and LTR scheduling model based on the image retrieval technology, obtained the unprecedented development. Firstly,we introduce an important application of visual features based on the depth of learning, the basic concept of visual similarity features and related generation algorithm is described in detail, including the optimization scheme of sparse matrix visual features. Then based on the generation of visual similarity characteristics of the whole architecture supports are introduced. And a detailed description of the architecture of the realization of technical points, and strategy algorithm. Finally, demonstrating the practical application of the visual similarity feature in image retrieval system and enhance the retrieval effect brought by th