Discriminative Dictionary Learning For Zero-Shot Image Classification

Zero-shot image classification(ZSIC)belongs to the field of image classification,which is a new technique inspired by the human beings' inferential ability.Its aim is to recognize the instances of unseen categories which have no training instances during training stage.ZSIC is typically achieved by transferring the knowledge from the seen classes to the unseen ones,the existing approaches bridge both the seen classes and unseen ones with one kind of class-level semantic representation,such as visual attributes and word vectors.In this way,how to connect the relations between the visual instances and the class-level semantic representations is a key challenge to ZSIC.Besides,the domain shift issue in ZSIC will degrade the classification performance since the training classes and testing ones are disjoint,therefore,solving the domain shift issue is another challenge to ZSIC.Firstly,this paper proposes a discriminative dictionary learning approach for zero-shot image classification.It reconstructs visual modality by using mapping modality of class semantic modality in a latent space,instead of directly constructing the relations between the visual modality and the class semantic modality with the labeled seen instances,which reduces the redundant information between class semantic modality of instances.Secondly,based on the aforementioned method,this paper further proposes a transductive discriminative dictionary learning approach for zero-shot image classification.In order to learn a more general mapping relation,the proposed method optimizes the learned model by using the predicted classification results of the test set,which ameliorates the domain shift issue in ZSIC.Finally,experimental results on three benchmark datasets(AwA,CUB and SUN)demonstrate the effectiveness and superiority of the proposed approaches to the state-of-the-art approaches.