基于深度特征的无监督图像检索研究综述

基于内容的图像检索(content-based image retrieval, CBIR)是一项极具挑战的计算机视觉任务.其目标是从数据库图像中找到和查询图像包含相同实例的图像.一个典型的图像检索流程包括2步:设法从图像中提取一个合适的图像的表示向量和对这些表示向量进行最近邻搜索以找到相似的图像.其中,决定图像检索算法性能的关键在于其提取的图像表示的好坏.图像检索中使用的图像表示经历了基于手工特征和基于深度特征两大时期,每个时期又有全局特征和局部特征2个阶段.由于手工特征的表示能力有限,近年来图像检索的研究主要集中在如何利用深度特征.将以提取图像表示的不同思路为线索,回顾无监督图像检索领域的发展历程,介绍该领域的一些代表性算法,并比较这些算法在常用数据集上的性能表现,最后探讨未来的研究方向.

A Survey on Unsupervised Image Retrieval Using Deep Features

Content-based image retrieval (CBIR) is a challenging task in computer vision. Its goal is to find images among the database images which contain the same instance as the query image. A typical image retrieval approach contains two steps: extract a proper representation vector from each raw image, and then retrieve via nearest neighbor search on those representations. The quality of the image representation vector extracted from raw image is the key factor to determine the overall performance of an image retrieval approach. Image retrieval have witnessed two developing stages, namely hand-craft feature based approaches and deep feature based approaches. Furthermore, there are two phases in each stage, i.e., one phase of using global feature and another phase of using local feature based approaches. Due to the limited representation power of hand-craft features, nowadays, the research focus of image retrieval has shifted to how to make the full utility of deep features. In this study, we give a brief review of the development progress of unsupervised image retrieval based on different ways to extract image representations. Several representative unsupervised image retrieval approaches are then introduced and compared on benchmark image retrieval datasets. At last, we discuss a few future research perspectives.