A Standardized Workflow for Illumination-Invariant Image Extraction

The illumination-invariant image is a useful intrinsic feature latent in colour image data. The idea in forming an illumination invariant is to postprocess input image data by forming a logarithm of a set of chromaticity coordinates, and then project the resulting 2-dimensional data in a direction orthogonal to a special direction, characteristic of each camera, that best describes the effect of lighting change. Lighting change is approximately simply a straight line in the log-chromaticity domain; thus, forming a greyscale projection orthogonal to this line generates an image which is approximately independent of the illuminant, at every pixel. One application has been to effectively remove shadows from images. But a problem, addressed here, is that the direction in which to project is camera-dependent. Moreover, preprocessing with a spectral sharpening transform to linearly transform the sensor curves to more narrowband ones greatly improves shadow attenuation, but sharpening is also camera-dependent and we may not have information on the camera. So here we take a simpler approach and assume that every input image consists of data in the standardized sRGB colour space. Previously, this assumption has led to the suggestion that the built-in mapping of sRGB to XYZ tristimulus values could be used by going on to sharpen the resulting XYZ and then seeking for an invariant. Instead, here we sharpen the sRGB directly and show that performance is substantially improved this way. This approach leads to a standardized sharpening matrix for any input image and a fixed projection angle as well. Results are shown to be satisfactory, without any knowledge of camera characteristics.