An approach to line labeling using Markov random fields

A method is presented for obtaining a line labeling from a greyscale image of opaque trihedral objects, using Markov random fields (MRFs). Orientation sensitive filters provide evidence to an MRF which marks the lines and junctions in the original image. These lines and junctions are then extracted and used to build a graph for another, coupled MRF, in which graph nodes represent individual junctions and lines rather than pixels. The paper emphasizes the design and operation of this latter, coupled MRF. An important aspect of this work is the fact that junctions and lines are both labeled by this coupled system, in contrast to earlier work, which took junction labels as given and immutable. This gives rise to a robust system, in which prior knowledge of acceptable configurations can serve to override noisy or misleading evidence