Residual-Prototype Generating Network for Generalized Zero-Shot Learning

Conventional zero-shot learning aims to train a classifier on training set (seen classes) recognize instances of novel classes (unseen by class-level semantic attributes. In generalized (GZSL), the needs both seen and unseen classes, which is problem extreme data imbalance. To solve this problem, feature generative methods have been proposed make up for lack classes. Current use class attributes as cues synthetic visual features, can be considered mapping attribute features. However, cannot effectively transfer knowledge learned from because information in features are asymmetric: contain key category description information, while consist that represented semantics. end, we propose residual-prototype-generating network (RPGN) GZSL extracts residual original an encoder–decoder synthesizes prototype associated with disentangle regressor. Experimental results show method achieves competitive four benchmark datasets significant gains.