Orientation and scale invariant symbol recognition using a hidden Markov model

The paper describes a method for symbol recognition based on encoding the boundary as a token sequence. Each token represents the local tangent to the boundary as a discrete region in a Hough space. A hidden Markov model was constructed for each symbol using a set of training examples. The Viterbi algorithm was then used to evaluate the probability of an unrecognised symbol being generated by each HMM. The method presented is translation, scale, and rotation invariant, and generates the orientation of the symbol relative to the training set exemplars.<
>