Physics-based Segmentation of Colour Images in Spherical Coordinates

We present image segmentation and highlight detection algorithms based on the dichromatic reflection model. For image segmentation, we use the model prediction that objects of a certain colour produce lines (the matte lines) radiating away from the origin of the RGB colour space. These lines therefore show up as peaks in a 2-dimensional histogram of the angular coordinates of a spherical polar coordinate representation of the RGB space. An algorithm for automatically locating these peaks is suggested. When the matte line locations are known, one can define cylindrical polar coordinate systems having their z-axes centred on the matte lines. We suggest a Hough-based algorithm for the detection of highlight lines in 2D-histograms of the ρ and z coordinates of the cylindrical polar coordinate system. Examples of the results of applying these algorithms are given. 1Updated April 22, 2004.