Knowledge-Based Multiple Lightweight Attribute Networks for Zero-Shot Learning

Zero-shot learning aims to recognize unseen objects. From the start of zero-shot learning research, attributes have played an important role. However, previous attribute-based methods do not fully exploit attributes and their relationships. In order to overcome these drawbacks, we propose a new framework. This framework consists of the convolutional neural network, fully connected networks and the attribute knowledge graph to make classification. This framework incorporates knowledge and is suitable for incremental learning. Also, the framework treats different attributes unequally according to their relationships, which further improves the recognition accuracy. Experiments show that this framework has achieved the comparable accuracy on the AWA2 dataset. At the same time, trainable parameters are reduced by about 100 times compared to previous attribute-based methods.