Research On The Pattern Recognition Methods Of Wood Surface Texture Based On GLCM

Texture is an important natural attribute of wood surface, which is directly related to the sense effect and economic benefit of woodware. The material performance and species of wood can be distinguished by it, and the wood physics takes it as a significant content of the woodiness environmental science to research. However, wood surface has fine complex structure, which is a hard problem puzzling wood academic circles and difficult to express with explicit mathematics analysis formula. And simultaneously the wood processing industry urgently needs the equipment to classify wood with their texture features. Consequently, to study wood surface texture has scientific and practical double value.In recent years, with the rapid development of the digital image processing technology and pattern recognition theory, the research of texture analysis and classification has made a series of breakthroughs. The research carried on wood surface texture classification and recognition by the digital image processing technology and pattern recognition theory, and the main content was as follows.Choose Betula-platyphylla, Pinus-koraiensis, Larix-gmelinii, Fraxinus-mandshurica, Quercus- mongolica as research objects, which are common in northeast, and establish the sample database including 1,000 samples of 10 texture categories.By analyzing the changing rule of Gray Co-occurrence Matrix (GLCM) feature parameters along with its three building factor including the making step d, the gray-level of image g and making directionθ, the way of building GLCM suitable to describe wood surface texture was established combined with its characteristics. The rule was that d equaled to 4, g equaled to 256, andθwas took as 0°, 45°, 90°and 135°four directions. The texture parameters value took the average of four directions to form rotation invariants.Based on the foregoing study, 14 texture parameters (W1～W14) of GLCM were obtained and their distribution in 10 texture categories was analyst at the same time.Three sets of wood surface texture parameter systems were founded by three different methods, which involved "the correlation analysis between parameters", "principal component analysis" and "the feature selection method based on simulated annealing arithmetic and the recognition rate of the nearest neighbor classifier". These texture parameter systems were as follows:①Wood surface texture parameter systemⅠ: Angular Second Moment (W1), Contrast (W2), Sum of Average (W6) and Sum of Variance (W7).②Wood surface texture parameter systemⅡ: PCAⅠ(Y1), PCAⅡ(Y2), PCAⅢ(Y3) and PCAⅣ(Y4).③Wood surface texture parameter systemⅢ: Angular Second Moment (W1), Variance (W5), Sum of Variance (W7), Inverse Difference Moment (W8), Variance of Difference (W9), Sum of Entropy (W10), Prominence of Clustering (W13).The classifier used in this study included the nearest neighbor (1-NN) classifier and the integrated BP neural network classifier. Under the three wood surface texture parameter systems above, the recognition rate of 1-NN classifier to unknown samples respectively were 85.25%, 86.75% and 87.50%, and the integrated BP neural network classifier were 86.50%, 87.00% and 90.25%. It showed that the recognition and classification capability of the integrated BP neural network classifier was better than the nearest neighbor classifier.Further analysis revealed that the dominant factor affecting the recognition and classification rate was to distinguish radial and tangential texture of the same wood species ineffctively, but Gauss-Markov Random Field was much good at it to GLCM. Therefore, 14 feature parameters of GLCM and 12 feature parameters of 5-rank GMRF were fused, and the forth feature parameter system was: V=[W2,W6,W7,W10, W13,θ1,θ3,θ5,θ9,θ10]. Under this texture parameter system, the recognition rate of the integrated BP neural network classifier to unknown wood samples gather was highly up to 97.00%, and the very satisfied results was gained.