Specular Free Spectral Imaging Using Orthogonal Subspace Projection

Specularity is an important issue in computer vision. Many algorithms have been proposed to remove highlights for color images. However, to our knowledge, no work has been done so far which specifically handles highlights in spectral imaging. In this paper, we introduce a specular invariant representation for hyperspectral images based on the dichromatic model and orthogonal subspace projection. It is a simple one step algorithm which only involves pixel-level operations, thus it does not require any segmentation. Nor does it require any pre/postprocessing or explicit spectral normalization. Importantly, unlike the previous methods for color images, it can be theoretically extended to handle highlights caused by multicolored illuminations. Experimental results demonstrate the effectiveness of our algorithm