Material based salient object detection from hyperspectral images

While salient object detection has been studied intensively by the computer vision and pattern recognition community, there are still great challenges in practical applications, especially when perceived objects have similar appearance such as intensity, color, and orientation, but different materials. Traditional methods do not provide good solution to this problem since they were mostly developed on color images and do not have the full capability in discriminating materials. More advanced technology and methodology are in demand to gain access to further information beyond human vision. In this paper, we extend the concept of salient object detection to material level based on hyperspectral imaging and present a material-based salient object detection method which can effectively distinguish objects with similar perceived color but different spectral responses. The proposed method first estimates the spatial distribution of different materials or endmembers using a hyperspectral unmixing approach. This step enables the calculation of a conspicuity map based on the global spatial variance of spectral responses. Then the multi-scale center-surround difference of local spectral features is calculated via spectral distance measures to generate local spectral conspicuity maps. These two types of conspicuity maps are fused for the final salient object detection. A new dataset of 45 hyperspectral images is introduced for experimental validation. The results show that our method outperforms several existing hyperspectral salient object detection approaches and the state-of-the-art methods proposed for RGB images.