Internal Defect Detection in Hardwood Logs With Fast Magnetic Resonance Imaging.

Abstract

Identification of defects such as knots in logs before the cutting operation would allow lumber mills to maximize the value of lumber from each log. This dissertation presented images obtained from scanning an oak log with magnetic resonance imaging (MRI). The unique characteristics of MRI images of hardwood logs were noted and were used to derive a quick algorithm to isolate defects. Defect regions had some pixels that varied considerably in intensity from their neighborhood, providing a seed for initiating the defect region. There was an overlap between the pixel gray level of the defects and clear wood. Therefore, traditional thresholding techniques did not cleanly separate these regions. In this study, region-growing methods were used to extract the defects. The algorithm grew the defect region seed until the border-pixel gray levels approached the average level of the neighborhood. The region-growing methods obtained more accurate defect regions than thresholding methods because of the simultaneous consideration of gray level and adjacency information. Two methods of MRI imaging were considered: spin-echo and echo-planar. Spin-echo imaging provided clear, detailed images but required about 20 seconds of acquisition time, which was too slow to be used in a production environment. Echo-planar images could be acquired in about 1/2 second, which was fast enough for production, but the images were fuzzy and noisy. The dissertation presented an algorithm that found the defect regions in spin-echo images. Region-growing methods use a number of parameters and the best parameters were unique for each image. However, common image statistics could be used to predict the proper parameters. The dissertation also presented an algorithm that found most of the defect regions in echo-planar images. Enhancing the echo-planar images using common general-purpose image-enhancement techniques failed because the lack of discrimination allowed the process to smooth image structures as well as noise. By taking advantage of the structure of a tree, smoothing between MRI frames accomplished the goal of smoothing along homogeneous areas and not across image structures. This "z-axis" smoothing enhanced the echo-planar image visually and reduced the number of false alarm defect regions.