Multiscale target extraction using a spectral saliency map for a hyperspectral image.

With the rapid growth of the capabilities for hyperspectral imagery acquisition, how to efficiently find the significant target in hyperspectral imagery has become a fundamental task for remote-sensing applications. Existing target extraction methods mainly separate targets from background with a threshold based on pixels and single-scale image information extraction. However, due to the high dimensional characteristics and the complex background of hyperspectral imagery, it is difficult to obtain good extraction results with existing methods. Saliency detection has been a promising topic because saliency features can quickly locate saliency regions from complex backgrounds. Considering the spatial and spectral characteristics of a hyperspectral image, a multiscale target extraction method using a spectral saliency map is proposed for a hyperspectral image, which includes: (1) a spectral saliency model is constructed for detecting spectral saliency map in a hyperspectral image; (2) focus of attention (FOA) as the seed point is competed in the spectral saliency map by the winner-take-all (WTA) network; (3) the multiscale image is segmented by region growing based on the minimum-heterogeneity rule after calculating the heterogeneity of the seed point with its surrounding pixels; (4) the salient target is detected and segmented under the constraint of the spectral saliency map. The experimental results show that the proposed method can effectively improve the accuracy of target extraction for hyperspectral images.