Shape recognition is one of the most familiar and fundamental problems in computer vision. And, it is important to recognize the shape regardless of its noise, size, position, orientation and starting point variation in the contour tracing.
Hidden Markov Model(HMM) is a formalized mathematical modeling tool and shows the excellent performance in the stochastic modeling of a system whose data are represented in ordered sequence i.e time-series. One of the limitations of the HMM is that it is needed time-series data with an explicit starting point. Therefore, it has been difficult to apply HMM to the shape recognition which has no explicit starting point while HMM have been successfully applied for the speech recognition and the writing signal recognition.
In this thesis, it is proposed a 2-D shape recognition model HYPROMI(HYpothesized process of Preparatory ROtation of a Mental Image ) based on the HMM and the theory of mental rotation, which is a physiological result about the human visual image processing. In HYPROMI, modified learning schemes consisted of "augmented-normal learning data" and "the slightly modified Baum-Welch reestimation algorithm" are used to overcome the explicit starting point needed in HMM
The most differentiable characteristics of the proposed model, HYPROMI, are as follows: HYPROMI is quite useful to the noise, scale, position, orientation, and starting point variation. HYPROMI can process any data sequence from an arbitrary starting point, therefore it does not need any preprocessing to find the reference point used to adjust orientation or starting point of a shape.
Experimental results show that HYPROMI is practical and effective in the shape recognition. As a result, this thesis gives a way of applying HMM for shape recognition such as the object recognition and off-line character recognition.