Automated Analysis of CT Images for the Inspection of Hardwood Logs

Abstract

This paper investigates several classifiers for labeling internal features of hardwood logs using computed tomography (CT) images. A primary motivation is to locate and classify internal defects so that an optimal cutting strategy can be chosen. Previous work has relied on combinations of low-level processing, image segmentation, autoregressive texture modeling, and knowledge-based analysis. Most previous work has also been limited to two-dimensional analysis of a single species only. This paper describes these approaches briefly, and compares them with a feed-forward neural-net classifier that we have developed. In order to accommodate species with different cell anatomies, CT density values are first normalized. Features are then extracted, primarily using local three-dimensional data. Somewhat surprisingly, this locality approach has resulted in a pixel-by-pixel classification accuracy of 95%. This accuracy improves during subsequent morphological processing steps which refine the detected defect regions in the images.

Citation

Li, Harbin; Abbott, A. Lynn; Schmoldt, Daniel L. 1996. Automated Analysis of CT Images for the Inspection of Hardwood Logs. Proceedings, IEEE International Conference on Neural Networks. 1744-1749.