Different Versions of Entropy Rate Superpixel Segmentation For Hyperspectral Image

This paper develops three new versions of engropy rate superpixel segmentation (ERS) for hyperspectral image superpixel segmentation by modifying the design of weights of ERS. The three new designed weights are based on Euclidean Distance (ED), Spectral Angle Distance (SAD) and spectral Correlation Coefficient (COR) respectively. Furthermore, two evaluation metrics are designed to analyze the performance of different versions of ERS. Experiment results on real hyperspectral images show that all proposed versions can obtain finely segmented superpixels. And compared with the commonly scheme which firstly obtain three principal components of the image before conducting ERS, both SAD and COR based ERS versions achieve better performance.