一种基于压缩Fisher向量的目标检索方法

摘要

The problem of object retrieval has been traditionally addressed with the bag of visual words (BoVW). But there are several problems existing in the conventional bag of visual words methods,such as the low time efficiency and large memory consumption. In this article,an object retrieval method based on compressed Fisher vectors is proposed for the above problems. Firstly,it is shown that show why the Fisher representation is well-suited to the object retrieval problem:it can automatically reduce the adverse effect of the object?from background. Then,Locality Sensitive Hashing(LSH)is used to compress Fisher vectors to reduce their memory footprint and computational costs and make it suitable for large scale database. Experimental results indicate that compressed Fisher vectors perform very well using as little as a few hundreds of bits per image,and significantly better than a very recent compressed BoVW approach.%当前，视觉词典法（BoVW，Bag of Visual Words）是解决目标检索问题的主要方法，但传统的BoVW方法具有词典生成时间效率低、检索内存消耗大等问题。针对这些问题，提出了基于压缩Fisher向量的目标检索方法，该方法首先将Fisher核机制用于目标检索，它能自动降低目标图像背景带来的不利影响，然后，采用位置敏感哈希（LSH，Locality Sensitive Hashing）对Fisher向量进行压缩编码以降低计算复杂度和内存开销，使之适用于大规模数据库。实验结果表明，新方法只用几百比特就能表征一幅图像内容，对大规模目标检索有很好的适用性，且较之当前主流的压缩视觉词典法具有更高的准确率。