基于锚图分类的在线半监督跨模态哈希

Online Semi-supervised Cross-modal Hashing Based on Anchor Graph Classification

近年来,哈希算法由于其存储成本小、检索速度快的特点,在大规模多媒体数据的高效跨模态检索中受到了广泛关注。现有的跨模态哈希算法大多是有监督和无监督方法,其中有监督方法通常能够获得更好的性能,但在实际应用中要求所有数据都被标记并不具有可行性。此外,这些方法大多数是离线方法,面对流数据的输入需要付出高额训练成本且十分低效。针对上述问题,提出了一种新的半监督跨模态哈希方法——在线半监督锚图跨模态哈希(Online Semi-supervised Anchor Graph Cross-modal Hashing, OSAGCH),构建了半监督锚图跨模态哈希模型,在只有部分数据有标签的情况下,利用正则化锚图预测数据标签,并通过子空间关系学习哈希函数,一步生成统一的哈希码,同时针对流数据输入的情况对该模型进行了在线化学习,使其能够处理流数据。在公共多模态数据集上进行了实验,结果表明所提方法的性能优于其他现有方法。

In recent years,hashing algorithm have been widely concerned in efficient cross-modal retrieval of large-scale multimedia data due to small storage costs and high retrieval speed.Most of the existing cross-modal hashing algorithms are supervised or unsupervised methods,and supervised methods usually achieve better performance.However,in real world applications,it is not feasible to require all data to be labeled.In addition,most of these methods are offline,which need to pay high training costs and are very inefficient when facing input of large stream data.This paper proposes a new semi-supervised cross-modal hashing me-thod--online semi-supervised anchor graph cross-modal hashing(OSAGCH),which builds a semi-supervised anchor graph cross-modal hashing model.It uses regularized anchor graphs to predict data labels in the case where only part of the data has labels,and uses subspace relationship learning to learn hash functions,generating a unified hash code by one step.Then the model is expanded to online version for streaming data input,allowing it to process streaming data.Experiments on public multi-modal data sets indicate that the performance of proposed method is superior to other existing methods.