Markov network modeling and recognition of closed contours

This thesis presents a method for pattern classification of two dimensional objects. The method is divided into three parts: preprocessing, inference, and classification. Preprocessing consists of the processing necessary to take a raw image containing an object and obtain a string encoding representing the object. Inference is the process of taking multiple related sample encodings and using them to train a constrained Markov network (CMN) using dynamic programming. Finally, classification is the process of aligning an unknown sample string encoding with multiple already trained CMNs so that the unknown string may be classified. In this alignment, the unknown string is considered cyclic, so a technique called channeling is incorporated to compute the alignment with less computational complexity than using a brute-force method of computing a full alignment for each possible offset. Experimentation of the method was performed using images of aircraft. Using various parameter values, statistical information concerning the inferred CMNs entropy and disagreement cost are presented and interpreted. Also presented, for various parameters settings, are statistics and interpretation of classification performance and mean rotational error of the cyclic alignment algorithm.