Invariant Image Improvement by sRGB Colour Space Sharpening

Reasoning from image formation, we have shown that there exists a greyscale image – the invariant image – that depends only on the reflectances in the scene. Since illumination dependence is removed, one aspect of the invariant image is that shadows are effectively removed. Moreover, given either a calibration, or clean data with good noise statistics, this invariant is easily found. However, we found that the performance was much poorer on ordinary images that include the typical nonlinear processing in cameras. The contribution of this paper is that we can find a good invariant notwithstanding input image nonlinearities. Our strategy is to follow standard colorimetric procedure and convert image RGBs to the appropriate colour space for our method. We do this by converting first to the linear sRGB colour space and then concatenating conversion to XYZ tristimulus values by a spectral sharpening transform. We handle a suite of images which were intractable to the original method and are now able to find a shadow-free intrinsic reflectance image.