Improved Feature Generating Networks for Zero-Shot Learning

Abstract

The purpose of zero-shot learning (ZSL) is to identify pictures of unseen classes, and there is no intersection between the training set and the test set. While the purpose of generalized zero-shot learning (GZSL) is to identify images from not only unseen classes but also seen classes, which is more challenging. Plenty of current ZSL and GZSL recognition methods are founded on feature generation methods like Variational Autoencoder (VAE), Generative Adversarial Networks (GAN) and so on to extenuate the data disproportion problem. While the previous works rarely focus on whether the features extracted from raw images have an impact on these generative models. In our work, we propose a novel architecture so as to ameliorate the original feature to enhance the performance of generating Networks not only for zero-shot Learning (ZSL) but also for generalized zero-shot learning (GZSL). Our approach utilizes a specific large-scale pre-trained model to gather the features from three distinct granular datasets namely AWA2, CUB, and SUN. Then, we concatenate the feature generated by this large-scale pre-trained model and the new feature of classification for training purposes generated by a dimensionality reduction method. Based on the state-of-the-art models, our means can raise the precision rates by about 15% for ZSL, and about 9.5% for GZSL, in all the conducted experiments on the three datasets. We further visualize the original feature and the feature processed by our method through t-SNE, and the result shows the feature data of the same class is more compact.