大规模图像检索中哈希排序方法研究

发布时间：2018-04-30 03:39

本文选题：自适应权重 + 哈希位加权　；参考：《吉林大学》2017年硕士论文

【摘要】：21世纪以来,随着各种数码拍照设备以及图像处理工具的不断普及,海量图像数据涌入人们的生活之中,现实中对图片检索、图像识别的需求使得在大型图像库中基于图像内容的检索变得非常重要。为了提高图像检索的效率,研究人员通过哈希方法将图像映射为二值哈希码,根据图像对应哈希码间的汉明距离来度量图像的相似程度。该方法具有压缩存储和快速查询的优势,不过仍有不足之处。当使用汉明距离排序时,哈希过程中的量化通常会使其辨别能力退化,与查询图像具有相同汉明距离的图像数量很多,这将导致不能有效地对这些图像进行排序,检索性能有限。因此,有必要对图像进行更细粒度的排序,哈希位加权已被证明是有效的解决方案。本文在前人对锚点图的研究基础上,利用数据点与锚点间的局部邻近关系矩阵来近似计算数据点间的原始相似性,克服了标准欧氏度量不能严格捕获数据点间全局相似性的缺点。该方法能够获得任何数据集中样本间相似性。本文基于查询点和数据库样本之间的相似性利用各哈希函数的辨别能力及它们对最近邻搜索能力的补充来学习一组查询自适应位权重,据此提出一种新的基于带权汉明距离的哈希排序方法。该方法将不同的权重分配给各个哈希位用以区分具有相同汉明距离的返回结果,从而获得更精细和准确的排序,这是一种能够用于所有种类哈希算法的通用加权方法,对数据分布没有严格的要求。最后根据本文提出的哈希排序方法在两个图像检索领域公认的数据集MNIST和NUS-WIDE上进行实验,表明本文所提出的方法比现有技术方法更具通用性和有效性。
[Abstract]:Since the 21st century, with the continuous popularity of various digital photography equipment and image processing tools, mass image data has poured into people's lives, and in reality, image retrieval. The requirement of image recognition makes the retrieval based on image content very important in large image database. In order to improve the efficiency of image retrieval, the researchers map the image to binary hash code by hash method, and measure the similarity of the image according to the hamming distance between the corresponding hash codes. This method has the advantages of compressed storage and fast query, but it still has some disadvantages. When the hamming distance is used in sorting, the quantization in the hash process usually degrades its discriminative ability. There are a lot of images with the same hamming distance as the query image, which will lead to the inability to sort these images effectively. Search performance is limited. Therefore, it is necessary to sort images more fine-grained. Hash weighting has proved to be an effective solution. In this paper, based on previous studies of anchor graph, the local proximity matrix between data points and anchors is used to approximate the original similarity between data points. It overcomes the shortcoming that standard Euclidean metric can not capture the global similarity between data points strictly. This method can obtain similarity between samples in any dataset. Based on the similarity between query points and database samples, this paper uses the discriminative ability of each hash function and their complement to the nearest neighbor search ability to learn a set of query adaptive bit weights. A new hash sorting method based on weighted hamming distance is proposed. In this method, different weights are assigned to each hash bit to distinguish the return results with the same hamming distance, so that more precise and accurate sorting can be obtained, which is a general weighting method that can be used in all kinds of hash algorithms. There is no strict requirement for data distribution. Finally, experiments are carried out on two recognized data sets, MNIST and NUS-WIDE, according to the hash sorting method proposed in this paper. The results show that the proposed method is more general and effective than the existing methods.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】