Discovering Unknown Labels for Multi-Label Image Classification

A multi-label learning (MLL) method can simultaneously process the instances with multiple labels, and many well-known methods have been proposed to solve various MLL-related problems. The existing MLL methods are mainly applied under the assumption of a fixed label set, i.e., the class labels are all observed for the training data. However, in many real-world applications, there may be some unknown labels outside of this set, especially for large-scale and complex datasets. In this paper, a multi-label classification model based on deep learning is proposed to discover the unknown labels for multi-label image classification. It can simultaneously predict known and unknown labels for unseen images. Besides, an attention mechanism is introduced into the model, where the attention maps of unknown labels can be used to observe the corresponding objects of an image and to get the semantic information of these unknown labels.