A New Approach to Automated Labeling of Internal Features of Hardwood Logs Using CT Images

Abstract

The feasibility of automatically identifying internal features of hardwood logs using CT imagery has been established previously. Features of primary interest are bark, knots, voids, decay, and clear wood. Our previous approach: filtered original CT images, applied histogram segmentation, grew volumes to extract 3-d regions, and applied a rule base, with Dempster-Schafer evidence accumulation, to identify 3-d objects. Initial feasibility results, however, were not entirely satisfactory because: (1) visual inspection of classification results indicates that the extent of features is not accurately circumscribed, (2) there is no quantitative estimate for the accuracy of classification, and (3) we strongly suspect that the speed of the previous approach would not lend itself easily to real-time implementation. Our current study has sought to improve on all of these limitations. In the current study, a pixel-level neural net classifier was trained using two hardwood species with vastly different cell anatomy. Ten-fold cross validation indicates that the classifier has a true error rate of approximately 9% at the pixel level. Subsequent morphological operations improve on this error rate. A third species, intermediate in cell anatomy to the other two, was also used to test the classifier. This feature labeling procedure, as compared to our previous approach, is faster, more accurate, less complex, and provides us with an estimate of classification accuracy.

Citation

Schmoldt, Daniel L.; Li, Pei; Abbott, A. Lynn. 1996. A New Approach to Automated Labeling of Internal Features of Hardwood Logs Using CT Images. Proceedings, Review of Progress in Quantitative Nondestructive Evaluation. 15: 1883-1890.