This thesis investigates the possibilities of using 3D X-ray tomography to detect and segment knots in logs. In the sawmill process this should be done previously to the sawing the log into boards. According to previous estimations, if comprehensive information about the knot geometry were available and the proper optimizations were implemented in the sawing process the sales value of the output could increase with up to 10%.

We assume that a 3D image of the interior of the log is acquired using two static X-ray sources and two corresponding 2D detectors, in between which the log is translated lengthwise at very high speed (max 3 m/s). To obtain reason­ able cost and fast read-out from the detectors we believe these should and could be assembled from a relatively small set of line-detectors (say, two times 10 or 20). Non-conventional reconstruction algorithms have been developed for the scanning geometry. However, due to both missing projection angles and relatively sparse 2D-detectors, the reconstructed volumes are far from perfect. Much of the knots is captured, while other features such as the boundary between the heartwood and sapwood is lost. Since the contrast between knots and heartwood is reasonably stable, while the contrast between knots and sapwood sometimes drops to zero, the general segmentation strategy is the following. Detection and segmentation begins in the heartwood from where we let a seed grow outwards into the sapwood, one knot at a time.

The use of second order derivatives to enhance elongated features such as knots has proven highly efficient in the present study. The derivative responses, generated in three different scales, are mapped rotation-invariantly into a new feature space, where we can discriminate between shapes of different second order variations (blobs, strings, planes, etc). Since knots are approximately string-like, voxels belonging to a knot can be distinguished from voxels belonging to reconstruction artefacts, which in our case appear in the form of planes. The result after such a shape discrimination process is a string-enhanced image, or rather volume. The segmentation consists of two parts, the first one dealing only with the heartwood section of the string-enhanced image. Here, we demonstrate that we can find the pith automatically, drawing heavily on the directional information embedded in the string data. The robust pith-finding procedure also leads naturally to a first segmentation in the heartwood. For the growth process into the sapwood we have implemented a fast marching process. The string values and their directional information are utilized in a navel algorithm where the seed is allowed to expand at a high speed only for high string-values and if the direction is in accordance with the assumed direction of the knot. The results are not totally convincing, however. Further experiments and optimizations of all the processing steps are needed to determine the viability of the whole approach.