Constrained Circular Hidden Markov Models for Recognizing Deformed Shapes

In this paper, we analyse the properties of the standard circular hidden Markov model (HMM) on 2D shape recognition. We point out the limitations of the circular HMMs and further propose to impose the constraint on the relationship between the initial and final states of circular HMMs to improve the performance. We develop two modified Viterbi algorithms to implement our proposal. The proposed algorithms have been tested on the database of the MPEG-7 Core Experiments Shape-1, Part B. The experiments show that both proposed algorithms can achieve better performance than that of the standard circular HMM in terms of accuracy. In particular, the second proposed algorithm, which is faster than elastic matching algorithms, has much potential due to its accuracy and speed.