EDGE-LABELING USING DICTIONARY-BASED RELAXATION

We present an improved application of probabilistic relaxation to edge-labeling. The improvement derives from the use of a representation of the edge-process that is internally consistent and which utilizes a more complex description of edge-structure. The particular novelty of the application lies in the use of a dictionary to represent permitted labelings of the entire context-conveying neighborhood of each pixel. This approach is to be contrasted with the use of approximate factorizations which have been employed in previous applications to decompose the neighborhood into object-pairs. We give details of the dictionary approach and the related representation of the edge-process. A comparison with other edge-postprocessing strategies is provided. This leads us to conclude that the dictionary-based approach is a powerful edge-postprocessing tool. It relaxes the demands on the level of filtering that has to be applied to cope with image noise with the benefit of reduced blurring of fine image features. © 1990 IEEE