Structural Alignment based Zero-shot Classification for Remote Sensing Scenes

Abstract

Zero-shot classification aims to classify unseen classes instances without any training data. However, the problem of class structure in consistency between visual space and semantic space severely affects zero-shot classification performance for remote sensing scenes. In order to tackle this problem, we employ semi-supervised Sammon embedding algorithm to modify semantic space prototypes to have a more consistent class structure with visual space prototypes. Then, unseen class prototypes in visual space can be effectively synthesized by transferring unseen knowledge from semantic space to visual space. Thus, classification task could be ultimately accomplished by the nearest neighbor method with the unseen class prototypes in visual space. The proposed method is extensively evaluated on two benchmark remote sensing scenes datasets, achieving the state-of-the-art performance.