Min-max correlation-based band reduction for edge detection in hyperspectral images

초록

This paper presents a statistical method of selecting significant bands from hyperspectral image for multi-dimensional edge detection. Due to the huge amount of data, the hyperspectral imagery may include information redundancy in highly correlated bands. The proposed method selects a set of significant bands in a recursive process by applying the min-max criterion to the pair-wise correlation coefficients among bands. The edge is then computed using the set of selected bands by considering each pixel as a spectral vector, and then edge magnitude in the direction of the maximum rate of change is computed. The experiments show that the result of edge detection is saturated with the number of bands and usage of all bands may not result in a meaningful edge image.