A De-redundant Network with Enhanced Classifier for Generalized Zero-Shot Learning

Jiayu Ding, Xiao Hu, Junjiang Xiang

Abstract: Generalized zero-shot learning (GZSL) is a challenging problem, which requires the model to classify both the categories that have appeared in the training set and those that have not. Existing generative models usually generate a large number of samples for unseen classes to tackle the serious class bias problem in GZSL. However, in the existing fine-grained data sets, the appearance of images in different categories exhibit slightly differences, which challenges the ability of models to generate discriminative examples and will certainly deteriorate the performance of GZSL. In order to tackle this problem, this paper integrates a de-redundant network (DR) and an enhanced classifier (EC) into a new model (DR-EC) to learn a feature space on which redundant information is removed and category information is enhanced. Specifically, the proposed model employs the de-redundant network to remove the redundant information from the original visual features and uses the enhanced classifier to make the non-redundant information in learned redundancy-free feature space more prominent. We evaluate our DR-EC model on three fine-grained benchmark datasets (i.e., CUB, SUN, and FLO) in GZSL. The results show that our DR-EC approach can outperform state-of-the-art methods with significant improvements.