A Semi-Supervised Learning Method for Hyperspectral-Image Open Set Classification

Abstract

We present a conceptually simple and flexible method for hyperspectral-image open set classification. Unlike previous methods, where the abundant unlabeled data inherent in the data set are ignored completely and unknown classes are inferred using score post-calibration, our approach makes the unlabeled data join in and help to train a simple and practical model for open set classification. The model is able to provide an explicit decision score for both unknown classes and each known class. The main idea of the proposed method is augmenting the original training set of K known classes using the pseudo-labeled unknown-category samples that are detected elaborately from the unlabeled data using modified OpenMax and semi-supervised iterative learning. Then a (K + 1)-class deep convolutional neural network model is trained based on the augmented training set with (K + 1) class samples. The model can not only classify instances of each known class but also refuse instances of unknown class explicitly. We validated the proposed method on four well-known hyperspectral-image data sets, obtaining superior performance over previous methods.