A Novel Hybrid Approach of Bayesian Theory and Neural Networks for Video Image Segmentation

Video image segmentation is essential for image analysis and the target recognition. In this study, a Bayesian theory and neural networks based image processing method was applied to video image segmentation. Firstly, a neural network with an incremental input node was designed for approximating to the posterior probability, which avoided the difficulty of estimation of class-conditional probability and could be applied to the occasions when prior probability changed. Secondly, the location information in the estimation of prior probability played a role in inhibiting the over-segmentation, and made the classifier more robust and flexible. Finally, a variable-step algorithm using the "Center of gravity" as the starting point for moving target diffused searching was developed. This algorithm could not only reduce noise, but also avoided the classification of each pixel in every video image, which facilitated to improve the performance of real-time.