Experiments with Som Based Inspection of Wood

We have devised a non-supervised clustering based approach for detecting and recognizing defects in lumber boards. The solution is simple to train, and supports detecting knots and other defects by using multidimensional feature vectors containing texture and color cues from small non-overlapping regions in the image. The key idea is to employ a Self-Organizing Map (SOM) for discriminating between sound wood and defects. An almost identical scheme is employed in classifying the defects. Human involvement needed for training is minimal. The approach is still under development, although it is approaching application level maturity. In this paper, we investigate the dependence between the size of the SOM and false alarm, error escape, and correct classification rates. Based on tests using demanding real-world material, rather small (12*8 nodes) SOMs provide attractive performance, approximately 31% false alarm and 5% error escape rate in defect detection. In defect classification, the accuracy is better than 72 % with the material used. All results are with respect to human region and defect labelings, and can be considered excellent. The approach is new in wood inspection, although we have tested a similar approach in food and steel inspection.