Deep Zero-Shot Image Classification Based On Adversarial Semantic Guidance

In the Internet era of big data,how to quickly and effectively classify and retrieve these data has become a hot topic in recent years.The traditional image classification technology usually needs to obtain a great number of labeled training samples related to the corresponding categories,and then extracts some of the samples to form training set to obtain the corresponding classifier.Finally,the classifier is employed to classify the test samples.However,the traditional image classification technology has certain limitations in reality.The acquirement of the annotation data requires expensive human and material cost.It is not realistic to collect the corresponding annotations for each category of image.To this end,the researchers proposed Zero-Shot Learning(ZSL),which deals with the image classification on unseen categories.This technique is utilized to transfer knowledge by learning the interrelationship between visual and semantics on seen classes to the unseen classes.Aiming at the task of current Zero-Shot Classification(ZSC),this thesis proposes a zero-shot classification algorithm based on Adaptive Weighted Hashing of Cycle-GAN(AWCH)and a triple discriminator adversarial network based on text reconstruction(Triple Discriminator GAN,TDGAN),respectively.Firstly,the research on hash-based learning is carried out for the problem of fast zero-shot classification and retrieval tasks.Specifically,both the visual and semantic modes are mapped into the new Hamming space and the distance metric is performed,hence the AWCH algorithm is proposed.The proposed method is a visual and semantic adaptive matching weight hash code,which maps the modal features into the Hamming space and fully exploits the related information between different modalities,thereby improving the utilization of cross-modal information.The proposed AWCH algorithm is applied to the zero-shot image classification and zero-shot image retrieval tasks.Compared with other ZSC approaches on the popular zero-shot dataset,the performance of the AWCH method is substantially improved.Secondly,the semantic gap between seen and unseen classes is currently common in ZSL.When the categories between seen classes and unseen ones are as unrelated as possible,the missing semantics of inter-class makes the transfer process particularly difficult.To solve this problem,the thesis further strengthens the fusion of zero-shot learning and adversarial learning,and deeply explores the more discriminative local semantic features to assist the ZSC.Therefore,we propose a triple discriminator adversarial network(TDGAN)algorithm based on text reconstruction which realizes the fusion of zero-shot learning and adversarial learning,and further explores the more discriminative local semantic features,so that images and text can be matched well,hence keeping semantic alignment.Extensive experiments on both CUB and NABirds datasets prove that our TDGAN consistently yields competitive performance compared to state-of-the-art ZSC approaches.