Research On Zero-shot Image Classification Based On Generative Adversarial Network

Image classification is the main research content in the field of computer vision.With the continuous development of deep learning,large-scale data is used by people to train different deep learning models.These models can achieve good results in image classification tasks.However,in practical applications,some image categories have a small number of samples,making it difficult to train a better model.And people hope that they do not have to retrain the model when they encounter new categories.In order to solve this problem,the concept of zero-shot learning was proposed,which meaning is that study the classification of train classes and test classes.At present,the zero-shot image classification task has made some progress,but it is difficult to establish a connection between the training class and the test class.To solve this problem,this paper uses the feature representation of the training category to generate the feature representation of the test category by generative adversarial network,and the zero-shot image classification task is converted into a classic image classification task.The main work of this article is as follows:(1)Zero-shot image classification based on feature generative adversarial network: This work uses generative adversarial network to solve the zero-shot image classification problem,respectively improve the generative network and discriminative network in generative adversarial network,and design FD-fGAN(Feature Discrimination Based on Feature Generative Adversarial Networks)model,which can solve the classic zero-shot image classification problem and the generalized zero-shot image classification problem.The experimental results show that the classification accuracy of the FD-fGAN model in multiple datasets is higher than the existed zero-shot image classification model.(2)Zero-shot image classification with attention mechanism: This work integrates the attention mechanism into the image feature generation process,and designs the FD-fGAN-Attention model to improve the quality of the generated image features.Therefore,a higher classification accuracy of zero-sample images is obtained.The experimental results show that when the difference between the categories is small,the model which incorporating the attention mechanism can further improve the performance of the zero-shot image classification task.