Classifying features in CT imagery: accuracy for some single- and multiple-species classifiers

Our current approach to automatically label features in CT images of hardwood logs classifies each pixel of an image individually. These feature classifiers use a back-propagation artificial neural network (ANN) and feature vectors that include a small, local neighborhood of pixels and the distance of the target pixel to the center of the log. Initially, this type of ANN was able to classify clear wood, bark, decay, knots, and voids in CT images of two species of oak (Quercus rubra, L., Quercus nigra, L.) with 95% pixel-wise accuracy. Recently we have investigated other ANN classifiers, comparing 2-D versus 3-D neighborhoods and speciesdependent (single species) versus species-independent (multiple species) classifiers using oak, yellow poplar (Liriodendron tulipifera, L.), and black cherry (Prunus serotina, L.) CT images. When considered individually, the resulting speciesdependent classifiers yield similar levels of accuracy (96-98%); however, all classifiers achieve greater than 91% accuracy. 3-D neighborhoods work better for multiple-species classifiers and 2-D is better for single-species. Multiple-species classifiers, whose training included both cherry and yellow poplar examples, exhibit the lowest accuracy. Nevertheless, when this combination of species is avoided, there is no statistical difference in accuracy between singleand multiple-species classifiers, suggesting that a multiple-species classifier can be applied broadly with high accuracy. Because all reported accuracy values are prior to postprocessing operations (which visually improve classification accuracy), we are confident that even the least accurate classifiers would be adequate for industrial implementation. Paper prepared for 3 IWSS 1998 Luleå University of Technology and IUFRO S5.04-10