A new HMM topology for shape recognition

This study deals with the shape recognition problem using the Hidden Markov Model (HMM). In many pattern recognition applications, selection of the size and topology of the HMM is mostly done by heuristics or using trial and error methods. It is well known that as the number of states and the non-zero state transition increases, the complexity of the HMM training and recognition algorithms increases exponentially. On the other hand, many studies indicate that increasing the size and non-zero state transition does not always yield better recognition rate. Therefore, designing the HMM topology and estimating the number of states for a specific problem is still an unsolved problem and requires initial investigation on the test data. This study addresses a specific class of recognition problems based on the boundary of shapes. The paper investigates the affect of the HMM topology on the recognition rate. A new topology, called circular HMM, is proposed and tested on the handwritten character recognition problem. The proposed topology is both ergodic and temporal. It eliminates the starting and ending states with the circular state transitions. The experiments indicate excellent performance compared to the classical temporal and ergodic HMM models.