A Feature Fusion Technique for Dimensionality Reduction

Merging features is necessary when it is advisable to use several different feature systems to solve applied problems. Such a problem arises, for example, in hyperspectral image classification, when the combination of spectral and spatial features significantly improves the quality of the solution. Likewise, several modalities can be used to identify a person, such as facial and hand features. The most commonly used feature merging method can be considered a simple concatenation. The problem with such a merger may be the different nature of the features, the need to use different dissimilarity measures, etc. To solve these problems, this paper proposes a feature fusion technique based on the transition to an intermediate form of data representation. A special case is considered when a space with the Euclidean metric is used as such a representation. The results of testing the proposed approach for hyperspectral data are presented.