编码选择哈希算法解决大规模图像检索问题

Bit selection hashing for large scale image retrieval

中文关键词  无监督哈希算法  编码选择  大规模图像检索

英文关键词  unsupervised hashing method  bit selection  large scale image retrieval

中文摘要

      哈希算法已被广泛应用于解决大规模图像检索的问题. 在已有的哈希算法中, 无监督哈希算法因为不需要数据库中图片的语义信息而被广泛应用. 平移不变核局部敏感哈希(SKLSH)算法就是一种较为代表性的无监督哈希算法.该算法随机的产生哈希函数, 并没有考虑所产生的哈希函数的具体检索效果. 因此, SKLSH算法可能产生一些检索效果表现较差的哈希函数. 在本文中, 提出了编码选择哈希算法(BSH). BSH算法根据SKLSH算法产生的哈希函数的具体检索效果来进行挑选. 挑选的标准主要根据哈希函数在3个方面的表现: 相似性符合度, 信息包含量, 和编码独立性. 然后,BSH算法还使用了一种基于贪心的选择方法来找到哈希函数的最优组合. BSH算法和其他代表性的哈希算法在两个真实图像库上进行了检索效果的对比实验. 实验结果表明, 相比于最初的SKLSH算法和其他哈希算法, BSH算法在检索准确度上有着明显的提高.

英文摘要

      Hashing methods have been widely used for solving large scale image retrieval problems. Among existing hashing methods, unsupervised hashing methods are most widely used because they do not require semantic information of images in the database. The shift-invariant kernelized locality-sensitive hashing (SKLSH) is a representative unsupervised hashing method which generates hash functions randomly without considering the performance of each projection. Therefore, weak hash functions yielding low retrieval performances may be generated by the SKLSH. In this work, we propose the bit selection hashing (BSH) which selects hash bits for the SKLSH based on the performance of hash bit projections in three aspects: similarity fitness, information capacity, and code independence. Then, a greedy selection method is applied to find the optimal combination of hash bits for the BSH. Two real world image databases are used to compare the performance of the proposed BSH with other representative hashing methods. Experimental results show that the BSH yields a significant improvement in comparison to the original SKLSH and other hashing methods.