Download (PDF 1.49 MB): https://research.fs.usda.gov/download/treesearch/21194.pdf
Abstract
This paper describes recent progress in the analysis of computed tomography (CT) images of hardwood logs. The long-term goal of the work is to develop a system that is capable of autonomous (or semiautonomous) detection of internal defects, so that log breakdown decisions can be optimized based on defect locations. The problem is difficult because wood exhibits large variations in texture along with irregular defect placement, particularly for hardwood species. In an earlier project, we developed a classification system that utilizes artificial neural networks (ANN) for this purpose. The system uses small neighborhoods in a CT image to make a preliminary classification decision for every pixel, using labels such as "knot," "split," and "bark." This approach has yielded high accuracy statistically. Subjectively, however, the results can often be improved through fiuther processing steps. For that purpose, we have developed a prototype system called IntelliPost, which can refine a segmented image. During its "learn mode," IntelliPost observes image-editing operations performed by a human operator, and develops its own rules based on those actions. Later, the user can place the system into "run mode" and provide new segmented images. The system automatically refines these new images by using the rules that it has developed. This approach allows IntelliPost to be tailored for different application domains (e.g., species and grading criteria) and for different user preferences. In tests involving CT datasets of red oak and sugar maple logs, the use of IntelliPost resulted in pixel-wise accuracy improvements ranging from 1.6 1 % to 19.47%.
Citation
Sarigul, Erol; Abbott, A. Lynn; Schmoldt, Daniel L.; Araman, Philip A. 2005. An interactive machine-learning approach for defect detection in computed tomogaraphy (CT) images of hardwood logs. In: Proceedings of Scan Tech 2005 International Conference, Las Vegas, Nevada, 15-26
