| As a basic task in the field of computer vision,image classification has been widely used in many real scenes.Using deep learning methods to solve image classification is a current research hotspot.However,when using the deep learning model for classification,especially in the process of supervised learning,its effect also benefits from large-scale labeled data.However,with the exponential growth of real-world data,the cost of manually labeling data is getting higher and higher,and some fields cannot even directly obtain a sufficient number of samples,which hinders real-world applications.For zero-shot image classification,only part of the labeled data is used in training,and the prediction of new categories during testing is expected to solve this problem.At present,the mainstream methods are divided into two categories: generative and embedded.To address the problems in both types of methods,the main research work of this paper is as follows:(1)A zero-shot image classification method based on enhanced VAEGAN(E-VAEGAN)is proposed,which mainly improves the generative model and solves the problem of generated sample quality.At present,the VAEGAN framework,which combines the generative adversarial networks and variational auto-encoder,has achieved good performance in zero-shot image classification.On the basis of VAEGAN,firstly,a feature alignment module is proposed.At the same time,the alignment center loss is introduced to achieve the alignment of image features and class attribute features.Then,the aligned features are fused with the hidden layer features of the encoder to improve the output features of the encoder.Second,triplet loss is applied during encoder training,further increasing the discriminative features.Finally,a transform module is proposed to feed back the hidden layer features of the discriminator to the generator to improve the quality of generated samples.(2)A generalized zero-shot image classification based on attribute-guided semantic learning is proposed,which mainly solves the problem of accurate location of attribute information in images and the learning of correlation between attribute information.This method consists of Associated Attribute Guidance Network(AAGNet)and Global Semantic Attention Network(GSANet).AAGNet uses graph convolution to learn the correlation between attribute text features,so as to accurately locate the local attribute information in the image,so as to obtain the attribute attention map representation of the image.GSANet further combines image features to obtain attribute activation features,completes the mapping from images to attributes,and then learns the global semantic relationship of attribute activation features through the Transformer encoder to generate accurate class attribute representations for classification.In this paper,the zero-shot image classification is studied,and the corresponding solutions are proposed for the problems of the two common methods.Finally,experiments on coarse-grained and fine-grained datasets show good classification performance compared to existing methods,and also validate the effectiveness and feasibility of the methods in this paper. |
