Research On Zero-shot Learning Based On Generative Model

Existing machine learning methods require a large amount of labeled data during training.However,considering that the training data of some classes is difficult to obtain or the acquisition cost is high in real life,and new classes are always dynamically added,it is difficult to obtain a large amount of available labeled data in this case.In order to train a good model without samples,researchers proposed the concept of zero-shot learning.At present,zero-shot learning methods mainly establish the connection between visible and unseen classes by embedding models,or train generative models to generate corresponding virtual samples for unseen classes.Generative model-based methods have attracted extensive research because they can generate virtual samples for unseen classes,thereby transforming zero-shot learning tasks into fully-supervised tasks.In the method based on the generative model,the final recognition accuracy of the model for the unseen category samples depends on the quality of the virtual samples generated by the model,so this paper takes this as the research focus to carry out the following work:(1)A zero-shot generation model learning method based on cyclic invariance is proposed.In the WGAN model,virtual samples are generated for invisible classes through a confrontation game between generators and discriminators.However,there are significant differences between the quality of the generated samples and the actual samples.Therefore,in this chapter,the corresponding intermediate samples will be calculated for the generated virtual samples to regularize the model generation;By mapping the generated visual features into semantic descriptions,the loss between the semantic description of the generated sample and the semantic description of the real category is calculated,and the model is optimized through the loss of cyclic consistency.(2)An attention based f-VAEGAN-D2 model method is proposed.In the f-VAEGAN-D2 model,virtual unseen class samples are generated by the VAE-GAN model,and two discriminators are used to identify the generated visible class samples and unseen class samples respectively.For selective attention to visual features,discriminative local features are emphasized to ensure the quality of samples generated by the model.In this method,a visual attention module is added to the model,and the visual features that have undergone selective attention are input into the discriminator to improve the discrimination ability of the discriminator,and then optimize the generator during the confrontation process to improve the quality of generated samples.(3)Based on the above research results,a zero-shot classification system based on generative models is designed and implemented.The system has designed a total of four modules,a visual display of data set selection,parameter debugging,and result display in zero-sample learning.In addition,a zero-sample image prediction system is designed in this paper,which can realize the classification of related images.determination.