2001 Volume 67 Issue 656 Pages 1092-1098
Bayesian estimation is often applied in pattern recognition problems. We formulate estimation errors of a posteriori Bayesian probabilities to be propagated from observation. Next, we apply the scheme of the formulation to a practical image recognition problem : based on a posteriori probabilities, sectionalized regions in outdoor-scene images are classified into five categories of landform elements, i, e., asphalt, concrete, sand/soil, gravel, and grass. The errors originate from RGB pixel values, and propagate to the a posteriori probabilities via intermediary HIS color measures within a region. We concretely clarify a mechanism of the propagation for all steps, and show an effectiveness of the scheme by adducing changeovers between a posteriori probabilities with two kinds of landform elements.