Research On Zero-shot Learning Based On Generative Model

In recent years,with the continuous improvement of my country’s scientific and technological level,various researches in the field of computer have made great progress,especially the image classification task in the field of artificial intelligence.The deep learning model is built and trained with the help of massive image data,and its classification accuracy has reached even surpasses human classification accuracy.However,in actual scenarios and applications,the massive data of some types of objects is difficult to obtain,or even almost zero,which has become an obstacle to the training process of traditional image classification models.In order to solve this problem,researchers proposed the concept of zero-shot learning in 2008,aiming to solve the image classification task when the number of samples of a certain category is very small or even missing.Learning to achieve the task of classifying images with missing samples.Since its proposal,zero-shot learning has gone through three development stages:attribute prediction,embedding model and generative model.The first two methods have made some progress in the field of zero-shot learning,but it is still difficult to establish a relationship between known categories and missing samples.The resulting zero-shot learning methods based on generative models,although the integration of generative models is in the To a certain extent,the problem of establishing connections between categories is alleviated,but the image samples generated by the existing zero-shot generation models generally have problems such as low quality,lack of details,and lack of practical significance.Aiming at the common problems of existing generative models,this paper proposes a zero-shot learning method based on generative models,which mainly includes two improvements:The first is to propose a zero-shot generation model Stack VAE-GAN based on variational autoencoders and stacked generative adversarial networks,aiming at the problem of unbalanced sample data between visible categories and unseen categories in existing zero-shot generation models.The GAN and VAE are connected through the parameter sharing of the generator and the decoding model,and the structure of the discriminant network is changed at the same time,so that it is necessary to discriminate whether the fake samples come from the variational autoencoder or the generative adversarial network while discriminating the authenticity of the samples.In order to learn the difference between the two,the image with higher resolution,more classification features and realistic meaning can be obtained.The algorithm combines two different generative models to achieve mutual learning and common progress,and completes the sample generation task in a staged structure to improve the model performance.Experiments are carried out on Aw A,CUB,FLO and SUN,and the accuracy of image classification is used as the evaluation standard.The experimental results verify the effectiveness of the algorithm.The second is to propose a zero-shot learning method based on word vectors and attention mechanism.In order to obtain samples with more diversity and classification features,based on the Stack VAE-GAN model,semantic word vectors containing more information are used to replace attributes.Information is used as prior knowledge to assist the generative model to complete the generation of samples,and at the same time,the attention mechanism is incorporated to help the model obtain samples with more effective features and improve the quality of sample generation.In addition,the attention mechanism is also incorporated into the feature extraction of visible class images,so that the classification model can learn more effective information to improve the performance of zero-shot image classification tasks.The classification test and related visualization experiments verify the effectiveness of the algorithm.