IDENTIFICATION OF LOG CHARACTERISTICS IN COMPUTED TOMOGRAPHY IMAGES USING BACK-PROPAGATION NEURAL NETWORKS WITH THE RESILIENT BACK-PROPAGATION TRAINING ALGORITHM AND TEXTURAL ANALYSIS: PRELIMINARY RESULTS

摘要

This research addressed the feasibility of identifying internal log characteristics in computed tomography (CT) images of sugar maple and black spruce logs by means of back-propagation (BP) neural networks with a resilient BP training algorithm. FiveCT images were randomly sampled from each log. Three of the images were used to develop the corresponding classifier, and the remaining two images were used for validation. The image features that were used in the classifier were gray-level values, textual, and distance features. The important part of the classifier topology, ie the hidden node number, was determined based on the performance indicators: overall accuracy, mean square error, training iteration number, and training time. For the training images, the classifiers produced class accuracies for heartwood, sapwood, bark, and knots of 99.3, 100, 96.7, and 97.9%, respectively, for the sugar maple log; and 99.7, 95.3, 98.4, and 93.2%, respectively, for the black spruce log. Overall accuracies were 98.5% for sugar maple and 96.6% for black spruce, respectively. High overall accuracies were also achieved with the validation images of both species. The results also suggest that using textural information as the inputs can improve the classificationaccuracy. Moreover, the resilient BP training algorithm made BP artificial neural networks converge faster compared with the steepest gradient descent with momentum algorithm. This study indicates that the developed BP neural networks may be applicableto identify the internal log characteristics in the CT images of sugar maple and black spruce logs.