A subspace regression and two phase label optimization for High Dimensional Image classification

Abstract

This paper introduces a two-step algorithm which deals with spectral mixing issue as well as performs the empirical study of continuous labelling in HSI images for over segmentation within class labels. In step-1, we have applied a subspace regression followed by an alpha expansion method to obtain the classified HSI image. This method better classifies the HSI image by removing the spectral mixing problem, which is a well-known problem in HSI domain. The classified image of step-1 is directly used in step-2 to improve the classification result by label update optimization using the energy of clusters. The optimization process in step-2 has performed in two phases. In the first phase of step-2, we have updated the cluster centers by the minimization of cluster energy. This energy minimization has stopped until some stopping criteria have met. The energy minimization has resulted in improved cluster centers. In the second phase of step-2, the RBF kernel based image function has updated using improved cluster centres, obtained from the phase-1. Classification probabilistic result from step-1 and updated image function from step-2 has transformed into a spectral data-cost. Subsequently, the data-cost of step-1 and step-2 have fused with the linear decision fusion method. Finally, The graph-cut method has applied to the fused spectral data-cost(Dc) and a spatial smoothness cost(Sc). Fusion of data-costs has resulted in a significant improvement in accuracy. © 2019, Springer Science+Business Media, LLC, part of Springer Nature.