基于特征聚类的轻量级图像搜索系统  

Lightweight Image Retrieval System Based on Feature Clustering

作　　者：王晓飞 周超 刘利刚[1] WANG Xiao-fei;ZHOU Chao;LIU Li-gang(School of Mathematical Sciences,University of Science and Technology of China,Hefei 230000,China;Tencent Computer Systems Co.,Ltd.,Shenzhen,Guangdong 518057,China)

机构地区：[1]中国科学技术大学数学科学学院,合肥230000 [2]腾讯计算机系统有限公司,广东深圳518057

出　　处：《计算机科学》2021年第2期148-152,共5页Computer Science

基　　金：国家自然科学基金(61672482)。

摘　　要：在图像搜索的场景中,由于搜索请求的随机性,为了提高搜索速度,搜索算法运行时需要把整个数据集预先载入到运行内存。由于运行内存价格远高于同容量的硬盘价格,降低运行内存自然可以大大降低图像搜索服务的成本,但如果直接对数据进行压缩,往往会极大地损失搜索精度。在这种情况下,文中提出了一种基于图像内容特征的分块式图像搜索框架。先利用神经网络的方法来预先提取图片特征,在不对特征进行量化压缩的前提下,采用一种启发式的聚类方法对数据进行分块,同时保证每个数据块的数据之间有一定的相似性。对于每个数据块,采用基于图结构的HNSW算法来构建索引子图以加速图片查询。在该框架下,通过控制查询时访问的数据块的个数,可以在保证精度的前提下大大减少算法所需要的运行内存容量。In the scene of image search,due to the randomness of search request,in order to increase the search speed,it is often necessary to preload the entire data set into the running memory.Because the price of running memory with the same capacity is much higher than that of hard disk,reducing the running memory can greatly reduce the cost of image search service.However,if the data is compressed directly,the search accuracy will be greatly reduced.In this case,this paper proposes a content-based ima-ge search framework,which divides data set into groups.Firstly,the neural network is used to extract image features.On the premise of not compressing the features,a heuristic clustering method is used to group the data,ensuring that there is a certain similarity between the data of each data group.For each data group,HNSW algorithm based on graph structure is used to construct index to speed up image query.In this framework,by controlling the number of data blocks accessed during query,the running memory capacity required by the algorithm can be greatly reduced,under the premise of ensuring the accuracy.

关 键 词：图像检索 相似搜索 聚类 图像特征提取 近似最近邻匹配 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]