Image sweep-and-mark algorithms. Part 2. Performance evaluations

The performances of three image preprocessing algorithms are evaluated to determine their potential for use in automated systems that identify defects in Douglas-fir (Pseudotsuga menziesii) veneer. The three algorithms are all sweep-and-mark algorithms, so called because after dividing the image into a regular rectangular array of areas called tiles, they sweep through the tiles gathering data on their features, then use this feature data to mark tiles that potentially contain defects. Subsequent algorithms can then concentrate on only the marked portion of the image, which usually constitutes only a small fraction of the original image data. A description of each algorithm was given in Part 1 of this paper (3). It is shown that for loose and tight knots, open holes, and pitch pockets/streaks, the statistical algorithm generally produced better results than either the morphological or color-cluster algorithms. It achieved an overall 99.0 percent accuracy in correctly identifying tiles containing knots and open holes, while maintaining an overall 85.5 percent accuracy in detecting clear wood. Assuming defects occupy 10 percent or less of the area of a typical sheet of veneer (a conservative estimate), this means that the image data can reliably be reduced by 77.0 percent, while ensuring a defect accuracy of 99.0 percent. The importance of this result is that the reduced image contains only essential information, thus reducing the data that subsequent algorithms must manipulate.

You must be logged in to download any documents. Please login (login accounts are free) or learn how to Become a Member