|Title:||MAP REPRESENTATIONS AND OPTIMAL ENCODING FOR IMAGE SEGMENTATION|
|Author(s):||KEELER, KENNETH CHARLES|
|Institution:||HARVARD UNIVERSITY; 0084|
|Source:||DAI, 51, no. 12B, (1990): 5907|
|Abstract:||In this thesis we consider an encoding-based approach to image segmentation, a problem in early machine vision in which one wishes to decompose the domain of a visual image into subsets on which the image data is homogeneous with respect to some property.
We begin with the study of a purely information-theoretic problem, that of designing length-efficient codes for information sources about which it is known only that the statistics obey some given partial order relationship. General results are obtained on the existence, behavior and efficiency of such universal codes. These results and a general coding-based philosophy are then applied to the image segmentation problem. We devise a new representation of segmented images, motivated by the desire to encode the information necessary for higher-level processing; this representation is as a labelled embedded planar map. To implement a Bayesian approach to estimating a segmented image from a given image we must assign a prior distribution (or, equivalently, a code length function) to segmented images, and in particular to the alphabet of maps. The probabilistic structure of these objects is not well-understood, but in the absence of such modelling there are reasonable order assumptions about their relative probabilities. Using the universal coding techniques mentioned above we drive upper and lower bounds on the length required to encode a planar map and the worst-case efficiency of this representation under these order assumptions. We then use this length as part of an estimation criterion for the segmented representation of an image. An algorithm for the approximate minimization of this criterion is described along with some experimental results.
ENGINEERING, ELECTRONICS AND ELECTRICAL