Deep Zero-Shot Hashing For Large-Scale Image Retrieval

Hash coding has been widely used in image retrieval.Recently,many deep hashing methods have been proposed and shown largely improved performance over traditional feature learning methods.Given a set of images with semantic annotations(such as category labels),hashing method tries to generate effective and compact binary codes.However,labeling large amounts of data is very intractable and cannot meet the real-time requirements.The existing hashing methods cannot obtain good retrieval results when the categories of images are not seen in the training dataset.How to transfer the hash function learned from the seen labeled dataset to the data of unknown categories(also known as zero-shot hashing),has become a hot issue for image retrieval.As far as we know,the existing zero-shot hashing methods are designed for single-label images,and there is no related research work on zero-shot hashing for multi-label image retrieval.Compared with single-label images,multi-label images have more abundant semantic information,which makes it more complex and difficult to study zero-shot hashing problem.The challenges come mainly from two aspects:(1)how to build the connection bridge to transfer the knowledge from the seen domain to the unseen domain;(2)how to solve the domain-shift between the seen and unseen classes.To solve these problems,this work innovatively proposes a transductive zero-shot deep hashing method(T-MLZSH).Specifically,we firstly build a visual-semantic bridge via instance-concept coherence ranking on the seen dataset with the help of word embedding learned from large-scale text libraries.Then we use this visual-semantic bridge to predict labels for the unseen dataset,which transfers the knowledge from the seen domain to the unseen domain.Next,we can train a deep hashing network on the seen labeled dataset and the unseen dataset with predicted labels.In addition,we design an instance-level pairwise similarity definition for multi-label image dataset to guide the hash model to learn more detailed similarity relations.Extensive evaluations on multiple multi-label datasets demonstrate that,the proposed hashing method achieves significantly better results than the compared methods.