Labeling Defects in CT Images of Hardwood Logs With Species-Dependent and Species-Independent Classifiers

This paper analyses computed tomography (CT) images of hardwood logs, with the goal of locating internal defects. The ability to detect and identify defects automatically is a critical component of efficiency improvements for future sawmills and veneer mills. This paper describes an approach in which 1) histogram equalization is used during preprocessing to normalize pixel values; 2) a feedforward neural network assigns tentative labels to individual image pixels; and 3) a morphological post-processing step removes noise and refines image regions. The normalization step facilitates the classification of wood features across different logs and different species. The neural network assigns tentative labels using normalized pixel values from small three-dimensional (3D) neighborhoods. We demonstrates the utility of this approach when the the network is trained using a single species of wood. This paper also considers the effect of training the network with samples from more than one species. Because small neighborhoods are used in either case, the classifier can be made to operate at real-time rates. Tests of the method using ten-fold cross-validation and CT images from three different logs resulted in a classification accuracy of approximately 95%.