COLOR AND TEXTURE BASED WOOD INSPECTION WITH NON-SUPERVISED CLUSTERING

The appearance of sawn timber has huge natural variations that a human inspector easily compensates for in his brain when determining the types of defects and the grade of each board. However, for automatic wood inspection systems these variations are a major source of complication. For instance, normal wood grain and knots should be reliably discriminated in all circumstances, but simple thresholding based detection methods used in state-of-the-art inspection systems frequently fail even in this apparently straightforward task. In this paper we compare color and texture features in defect detection from lumber. A non-supervised clustering based approach is used for discriminating defects and sound wood. The solution is simple to train, and supports detecting knots and other defects by using multidimensional feature vectors containing texture and color cues for small non-overlapping regions in the image. A key idea is to employ a Self-Organizing Map (SOM) for discriminating between sound wood and defects. The experiments show that color histogram features together with local binary pattern based texture measures are a very promising method. The non-supervised training approach contributes to low error escape and false alarm rates.