Research On Zero-shot Image Classification Based On Attribute Learning

The attribute serves as a bridge between underlying features and category labels to realize knowledge transfer between seen and unseen classes,which provides an effective solution to the deficiency of category labels in zero-shot image classification(zero-shot learning).However,zero-shot learning still faces the problem that the classifier trained on seen images is difficult to test unseen images directly.Therefore,this paper focuses on the zero-shot image classification problem based on attribute learning.The main work includes:On one hand,during generating features by using the attribute as a control condition in generative adversarial network,due to the lack of generated features and attributes,as well as constraints on generated and real features,it is hard to fully characterize the relationship between generated features and corresponding attributes or real features.For the above problems,the zero-shot image classification model based on double-cycle loss generated features is proposed,which consists of two generative adversarial networks and a variational auto-encoder.First,the first generative adversarial network generates virtual features based on original attributes.Second,the variational auto-encoder is used to perform attribute reconstruction by generated virtual features.The constraint between original attributes and reconstructed attributes through attribute cycle loss ensures that generated virtual features can express semantic attributes better.Then,the second generative adversarial network regenerates the features based on reconstructed attributes.The constraint between real and generated features through the feature cycle loss ensures the generated virtual features are closer to real features.Finally,the Softmax classifier is jointly trained with seen class features and generated virtual features of unseen classes to predict class labels for test images.On the other hand,for categories with similar attribute descriptions,the generative adversarial network uses the attribute as a control condition to generate features without considering differences between attributes of different categories.Meanwhile,it merely relies on attributes to generate features and ignores the influence of sample features on generated results.To this end,the zero-shot image classification model based on hybrid attributes and generated features under the transductive zero-shot setting is proposed.First,image features are reconstructed using the variational auto-encoder to obtain latent features,which contain image characteristic information in the latent space.The latent features and original attributes combine to form hybrid attributes.Then,using hybrid attributes as a control condition,the generative adversarial network generates virtual features.Hybrid attributes increase the difference of semantic descriptions between categories with similar attributes to generate more real virtual features.Finally,seen class features and virtual features of unseen classes generated by hybrid attributes are jointly used to train the Softmax classifier to predict labels of test images.Experimental results on CUB,FLO,SUN and AWA datasets verify the effectiveness of the zero-shot image classification models proposed in this paper.