Automated labeling of log features in CT imagery of multiple hardwood species

Before noninvasive scanning, e.g., computed tomography (CT), becomes feasib le in industrial sawmill operations, we need a procedure that can automatic ally interpret scan information in order to provide the saw operator with i nformation necessary to make proper sawing decisions. To this end, we have worked to develop an approach for automatic analysis of CT images of hardwo od logs. Our current approach classifies each pixel individually using a fe ed-forward artifical 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 ANN was able to classify clear wo od, bark, decay, knots, and voids in CT images of two species of oak with 9 5% pixel-wise accuracy. Recently we have investigated other AMN classifiers , comparing 2-D versus 3-D neighborhoods and species-dependent (single spec ies) versus species-independent (multiple species) classifiers using oak (Q uercus rubra L. and Q. nigra L.), yellow-poplar (Liriodendron tulipifera L. ), and black cherry (Prumus serotina Ehrh.) CT images. When considered indi vidually, the resulting species-dependent classifiers yield similar levels of accuracy (96-98%). 3-D neighborhoods work better for multiple-species cl assifiers, and 2-D is better for the single-species case. Classifiers combi ning yellow-poplar and cherry data misclassify many pixels belonging to spl its as clear wood, resulting in lower classification rates. If yellow-popla r was not paired with cherry, however, we found no statistical difference i n accuracy between the single-and multiple-species classifiers.