Distance-based hyperspectral open-set classification of deep neural networks

Using deep learning, the supervised hyperspectral image (HSI) classification methods are based on the ideal assumption of closed sets, in which all testing classes are defined a priori. However, it is impossible to collect all classes while training in the open set setting where unknown classes can be submitted while testing. Traditional deep neural networks lack the ability to perceive the unknown classes without external supervision; therefore, they will misclassify them as known classes. This problem is called the open set classification (OSC). This paper proposes a distance-based OSC method for deep neural networks to solve the problem. The method introduces the class anchor clustering (CAC) classifier and loss to train deep neural networks, which can make known classes cluster around fixed class centres in the logit space. Boxplot and extreme value theory (EVT) are then used to determine the distance between input instances and their nearest class centres for unknown rejection without sacrificing the classification accuracy of known classes dramatically. Many experiments on HSI datasets indicated the validity of the HSI-OSC method.