An Image Retrieval Algorithm Via Hash Code Learning And Zero-shot Learning

The rapid development of advanced techniques,e.g.,prevalent smart mobile device,fast-evolving social media platform,cheap mass storage,launches the big multimedia era,remarkably reshaping the way people sharing,self-expressing and communicating.People produce a lot of data in internet life every day,especially multimedia data such as picture and video.In the big data era,how to retirev and save image efficiently has attracted wide attention.Essentially,hashing transforms high-dimensional data into compact binary codes,which supply encouraging solutions in saving storage and computational cost.Generally,existing hashing approaches can be partitioned into two groups: data-independent and data-dependent.Based on the level of supervision,the data-dependent methods can be further categorized into two subgroups: unsupervised and supervised.It has been extensively demonstrated that supervised hashing can achieve satisfactory performance because of using image feature and label.Admittedly,supervised hashing is able to achieve state of the art performance in many application scenarios,I argue that it may probably fail under the circumstance of the rapid evolution of newly-emerging concepts and the corresponding multimedia data.On the one hand,due to the expensive cost of gather manually labeled training,existing systems may suffer from scarcity of training data of “unseen” categories.On the other hand,the unaffordable cost of re-training decreases the flexibility and reusability of existing hash approaches.Meanwhile,supervised hashing inevitably suffers from the fundamental issue of semantic gap,i.e.,the difference between low-level representation(e.g.,original visual feature,binary codes)and high-level human cognition.In order to conquer the aforementioned obstacles,in this paper,we propose a novel approach,termed Attribute Hashing(AH),which enables efficient and effective image retrieval in zero-shot scenario.We can construct the hashing model in the training time by use of seen categories and then unseen category can leverage the well established mapping from the relevant seen categories through attribute.AH avoids scarcity of training data and cost of re-training,which brings hope for image retrieval by hashing;AH build a multi-layer hashing hierarchy to model the relationships among visual feature,binary codes,attributes and labels.AH can narrow the semantic gap between binary codes and labels through the multi-layer hierarchy.In order to further guarantee the quality of binary codes and hash functions,AH can preserve the discrete nature and intrinsic local structural information among data.The discrete optimization strategy can avoid the accumulated errors caused by relaxation in traditional continuous optimization methods.Extensive experiments on several real-world image datasets show the superiority of AH approach as compared to the state-of-the-arts.