Image segmentation for defect detection on veneer surfaces.

摘要

Machine vision is widely used in scientific areas and non-wood using industries, but the extreme variability of wood has limited its adoption by forest products industries. However, it is now becoming a key factor in further automation of the forest products industry. As a very important part of machine vision, developing image segmentation algorithms that can be used for wood products is an ambitious undertaking. The focus of this research was to adapt existing and develop some new segmentation algorithms which could be used to detect defects on veneer surfaces.; Nine algorithms covering three segmentation technique categories were explored. Three existing edge detection algorithms were modified for use on veneer images, and four existing thresholding algorithms were adapted in both global and local versions. Two new region extraction algorithms were developed specifically for defect detection on veneer surfaces.; The performances of these nine algorithms were tested and compared under the combinations of two camera resolutions (5-bit and 8-bit), three color spaces (RGB, Lab, and gray-scale), and seven surface features (clear wood, blue stain, loose knot, pitch pocket, pitch streak, tight knot, and wane). Ten sample images for each of seven surface features on Douglas-fir veneer (Pseudotsuga menziesii) were used. Ten measures were proposed for performance evaluation. A multi-factor factorial ANOVA was used in the performance tests and comparisons.; The best combinations of camera resolution and color space for each of the algorithms were determined. The 5-bit and 8-bit camera resolutions were not significantly different for the three edge detection and two region extraction algorithms, but the 8-bit camera resolution was better for all but one of the thresholding algorithms. That exception was the global Otsu thresholding algorithm, for which the 5-bit camera resolution was better. The RGB color space was the best for all algorithms. Overall, the two region extraction algorithms were the best. Under the best combination of factors, those two algorithms provided the highest defect detection accuracies of 91% for pitch streak samples and over 95% for loose knot, tight knot, and pitch pocket samples. These results were accomplished while still providing clear wood accuracies of over 95%. The one performance exception was blue stain, for which no satisfactory algorithm was found.