The lumber industry has given special attention for lumber grading and selection stages. Machine Vision Systems have been proposed as a technological solution for automation of these stages. The proper feature selection for discriminating defect and clear wood is one of the most challenging in the development of such technology. The objective of this work was to evaluate, using multivariate analysis, the discriminating power of color images percents. In this work, linear and quadratic discriminant analysis were accomplished for classification of defects and clear wood in digital images of eucalyptus lumber. The percent features of the histogram for the red, green and blue bands, from two sizes of image blocks were used for developing and testing the discriminant functions. 492 blocks were used, containing the 12 studied defects and clear wood, derived from images of 40 lumbers randomly sampled. The features were analyzed with their original values, scores of the principal components and scores of the canonical variables. The smallest global misclassification errors were 19% and 24% for linear discriminant function with the canonical variable scores using block sizes of 64x64 and 32x32 pixels, respectively. The percent features were considered appropriate to discriminate defects and clear wood in digital images.
Image processing; pattern recognition; eucalyptus lumber grading
