Spectral Generalized Category Discovery by training on combined labels

Hyperspectral imaging poses major challenges for supervised classification methods due to the high dimensionality of the data. This classification was solved in previous works with known classes. In this paper we present a different aim where the goal is to discover and distinguish novel categories for hyperspectral images in scenarios where some classes are unknown. This belongs to Generalized Category Discovery (GCD) problem with a further task for automatic clustering of unlabeled data with partial knowledge. Our contribution is an additional supervised learning stage on combined labels coming from two sources. Our approach consists of three stages, (i) dimension reduction, (ii) clusters assignment with the Hungarian method, and (iii) classification, where a special training set is constructed from ground true labels and the filtered test set with (dummy) labels that are the predicted categories from the clustering part. The proposed method (S-GCD) was evaluated on two datasets, and our experiments demonstrate that classification on combined labels greatly improves both classification and clustering scores.