This thesis describes the design of an image interpretation system for the automatic detection of internal hardwood log defects. The goal of the research is that such a system should not only be able to identify and locate internal defects of hardwood logs using computed tomography (CT) imagery, but also should be able to accommodate more than one type of wood, and should show potential for real-time industrial implementation. This thesis describes a new image classification system that utilizes a feed forward artificial neural network as the image classifier. The classifier was trained with back-propagation, using training samples collected from two different types of hardwood logs, red oak and water oak. Pre-processing and post-processing are performed to increase the system classification performance, and to make the system be able to accommodate more than one wood type. It is shown in this thesis that such a neural-net based approach can yield a high classification accuracy, and it shows a high potential for parallelism. Several possible design alternatives and comparisons are also addressed in the thesis. The final image interpretation system has been successfully tested, exhibiting a classification accuracy of 95% with test images from four hardwood logs. / Master of Science

Log sawing simulation computer programs can be a valuable tool for training sawyers as well as for testing different sawing patterns. Most available simulation programs rely on databases from which to draw logs and can be very costly and time-consuming to develop. In this study, a computer program was developed that can accurately generate random, artificial logs and serve as an alternative to using a log database. One major advantage of using such a program is that every log generated is unique, whereas a database is finite. Real log and external defect data was obtained from the Forest Service Northeastern Research Station in Princeton, West Virginia for red oak (Quercus rubra, L.) logs. These data were analyzed to determine distributions for log and external defect attributes, and the information was used in the program to assure realistic log generation. An attempt was made to relate the external defect attributes to internal defect characteristics such as volume, depth, and angle. CT scanning was used to obtain internal information for the five most common defect types according to the Princeton log data. Results indicate that external indicators have the potential to be good predictors for internal defect volume. Tests performed to determine whether a significant amount of variation in volume was explained by the predictor variables proved significant for all defect types. Corresponding R2 values ranged from 0.39 to 0.93. External indicators contributed little to the explanation of variation in the other dependent variables. Additional predictor variables should be tested to determine if further variation could be explained. / Master of Science