Parameterization methodology for 2D shape classification by hidden Markov models

In computer vision, two-dimensional shape classification is a complex and well known topic, often basic for three-dimensional object recognition. Among different classification methods, this paper is focus on those that describe the 2D shape by means of a sequence of d-dimensional vectors which feeds a left to right hidden Markov model (HMM) recogniser. We propose a methodology for featuring the 2D shape with a sequence of vectors that take advantage of the HMM ability to spot the times when the infrequent vectors of the input sequence of vectors occur. This propierty is deduced by the repetition of the same HMM state during the moments in which the infrequent vectors is repeated. These HMM states are called by us synchronism states. The synchronization between the HMM and the input sequence of vectors can be improved thanks to adding an index component to the vectors. We show the recognition rate improvement of our proposal on selected applications.