Multi-Scale Attribute-Based Attention For Fine-Grained Zero-Shot Learning

Zero-shot recognition achieves the classification of unseen classes by aligning the visual features and semantic information of images, while the global features extracted by zero-shot learning method are not enough to classify fine-grained classes. Firmed at the foregoing shortages, we put forward a multi-scale attribute-based attention network that captures local feature differences between fine-grained categories and improves the extraction efficiency of different scale attributes. Instead of aligning global visual features with the associated class semantic features, we achieve this by embedding each attribute which aims to pay more attention to the regions most related to the attribute. To this end, we adjust the contribution to prediction of each attribute by using multi-layered attention mechanism. Moreover, we design a new joint calibration loss to solve the distribution deviation problem caused by ignoring the clustering between coarse-grained classes in the fine-grained classification process. We perform experiments on several popular classical datasets of CUB, SUN and AWA2, demonstrating that our proposed framework works better than existing methods.