基于局部特征聚合的图像检索方法

Image Retrieval Based on Locally Features Aggregating

多媒体数据尤其是图像数据的急剧增长,使得基于图像内容的检索成为一个非常重要的研究课题.图像的特征描述以及特征的索引机制是实现基于内容图像检索的关键.针对图像局部聚合描述符(Vectors of LocallyAggregated Descriptors,VLAD)中硬分配难以准确描述局部特征向量与聚类之间隶属关系的问题,采用软分配策略,根据局部特征向量与聚类中心的距离分配不同的隶属权值,生成更具代表性的软分配局部聚合描述符(SoftAssignment-VLAD,SA-VLAD).针对非对称距离计算倒排索引机制(Inverted File with Asymmetric Distance Com-putation,IVFADC)在查询时为保证结果的查全率而增加候选倒排索引链的数目,导致距离计算和查询时间增加的问题,提出引入简单的散分配方法,将可能落入多条链表中的数据库向量进行多次编码,实现了基于散分配的非对称距离计算倒排索引机制(Dispersed Assignment-IVFADC,DA-IVFADC).实验结果表明:DA-IVFADC机制与SA-VLAD描述符,在很大程度上减少了查询时间,同时有效提高了查询结果的准确率.

With the amount of multimedia data,especially image data,increased rapidly,similar image retrieval has become a very important research subject.Features which describe images and feature indexing are important to image retrieval.Vectors of Locally Aggregated Descriptors（VLAD） cannot describe the relationship between local descriptors and clusters.More representative aggregation vector called Soft Assignment-Vectors of Locally Aggregated Descriptors（SA-VLAD） is put forward by using soft-assignment.Soft-assignment is a distributive strategy that the weight assigned to neighboring cells depends on the distance between the local descriptors and the cell centers.Besides,based on Inverted File with Asymmetric Distance Computation（IVFADC）,a new indexing scheme named Dispersed Assignment-Inverted File with Asymmetric Distance Computation（DA-IVFADC） is implemented by using dispersed-assignment in indexing stage to resolve the problem of massive distance computation.Experimental results demonstrate that DA-IVFADC and SA-VLAD lessens the query time to a great extent and effectively improves the accuracy rate of results.