A new multiscale Bayesian model averaging framework for texture segmentation

In texture segmentation, in order to accurately classify any pixel a suitable neighborhood must be chosen. However, selecting a neighborhood size and orientation is a difficult and often ad hoc task. We view the task of choosing a neighborhood as model selection problem and develop a multiscale Bayesian model averaging (BMA) framework for pixel-level texture segmentation. This framework leads to a maximum a posteriori (MAP) segmentation rule that combines information from different neighborhoods (models) defined at multiple scales and locations. Thus, our new method avoids the unsatisfactory requirement of a user-specified notion of “neighborhood,” instead letting the data speak for themselves. The performance of the new segmentation algorithm is examined in simulated studies