基于加权的多流形排序的图像检索研究

发布时间：2018-03-15 23:17

本文选题：基于内容的图像检索　切入点：特征提取与表示　出处：《东北师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：计算机网络技术、多媒体技术不断发展,以及摄录设备、存储设备快速更新,图像成为信息的主要载体之一。随着大数据时代到来,在海量图像数据中有效地查找需要的信息,是人工智能与机器学习领域一个热门研究问题。90年代后期,逐步出现了基于内容的图像检索(Content-based Image Retrieval,简称CBIR)技术,能够借助算法对图像的内容属性及语义进行分析,客观、全面地描述并搜索图像,在网络图像搜索、医学图像分析、安全监控、不良信息过滤和电子商务等领域得到了广泛的应用。目前,基于内容的图像检索技术得到众多科研机构及学者的普遍关注,许多经典的检索方法和成功的检索系统相继提出。但检索问题具有的复杂性、结构性的特点,对图像的特征提取与表示、相似性度量的研究是这些方法的关键问题。为了获得更高的检索性能,本文提出基于加权的多流形排序算法,并应用于基于内容的检索任务,主要工作如下:(1)采用了多视角提取图像特征。图像内容类内差异较大,特别是自然图像存在环境变化大及噪声干扰现象。本文同时采用颜色直方图、边缘直方图和局部二值模式分别描述图像内容的颜色、形状和纹理信息,多角度描述了丰富多样的图像内容;(2)提出了多视角快速构建图结构,探索了数据的结构信息,保障了多流形排序的有效进行;(3)探索了多种加权策略融合多流形排序。多流形排序利用最大值融合、平均值融合和线性加权融合对多视角下单流形排序进行融合,能够挖掘不同视角间的相关性。与经典检索方法不同,基于加权的多流形排序的检索方法在特征提取与表示、相似性度量上均有所贡献。在Corel1000数据集、Oliva数据集和Caltech5138数据集进行检索实验时,实验结果验证了基于加权多流形排序的图像检索方法的有效性,特别是基于线性加权融合的多流形排序算法能够利用视角间的互补信息对样本得分进行估计,达到高效检索的目的。
[Abstract]:With the continuous development of computer network technology and multimedia technology, as well as the rapid updating of recording equipment and storage equipment, images become one of the main carriers of information. It is a hot research problem in the field of artificial intelligence and machine learning. In the late 1990s, Content-based Image Retrieval (CBIR) technology appeared gradually, which can analyze the content attributes and semantics of images with the help of algorithms. It has been widely used in the fields of network image search, medical image analysis, security monitoring, bad information filtering and electronic commerce. Content-Based Image Retrieval (CBIR) technology has been widely concerned by many scientific research institutions and scholars. Many classical retrieval methods and successful retrieval systems have been put forward one after another. However, the retrieval problem has the characteristics of complexity and structure. Similarity measurement is the key problem in image feature extraction and representation. In order to achieve better retrieval performance, a weighted multi-manifold sorting algorithm is proposed and applied to content-based retrieval. The main work of this paper is as follows: (1) the image features are extracted from multiple perspectives. There are great differences in the image content, especially in the natural images with large environmental changes and noise interference. The color histogram is also used in this paper. Edge histogram and local binary pattern describe the color, shape and texture information of image content respectively. In this paper, multiple weighted strategies are explored to fuse multi-manifold sorting. Multi-manifold sorting uses maximum fusion, mean value fusion and linear weighted fusion to merge multi-view order manifold ordering. Different from the classical retrieval methods, the weighted multi-manifold sorting retrieval method is based on feature extraction and representation. In the experiments of Corel1000 dataset and Caltech5138 dataset, the effectiveness of the image retrieval method based on weighted multi-manifold sorting is verified. In particular, the multi-manifold sorting algorithm based on linear weighted fusion can estimate the sample score by using complementary information from different angles of view to achieve the purpose of efficient retrieval.
【学位授予单位】：东北师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】