Research For Wood Classification Based On SEM Micrographs

Advanced computer image processing and machine learning techniques have beenapplied to feature extraction and wood classification based on wood tansverse-sectionscanning electron microscopy (SEM) micrographs in this study. Firstly, the imagewould be partioned into multiple segments by ThreshCanny and Graph Cuts beforeextracting the features reflecting the wood microstructure. Then, the dimensions of thefeatures were reduced by principal component analysis. After feature fusion, these woodstructural features were combined with the machine learning techniques to explore thefeasibility of wood classification.The wood feature parameters proposed in this study, which based on ThreshCannyand Graph Cuts algorithm have a good distinguishing ability for the wood SEMmicroscopic images. In a given wood species set, the feature parameters of theearlywood and latewood in the same kind of softwood, and the feature parameters indifferent kinds of softwood, both all existing significantly differences. Moreover,comparing with softwood, the differeces between the feature parameters in the samekind of hardwood was not obvious, and it was very difficult to find the obviousboundary value between some feature parameters. However, obvious differencesbetween the feature parameters in different kinds of hardwood still existed.Ten feature parameters about the SEM micrograph of wood transverse-sectionwere analyzed by principal component analysis. The experiment results show that thecontribution of first two principal components to total feature parameters was closed to85%. Meanwhile, the interdependency among the feature parameters was analyzed toextract6features which were relatively independent and significant for woodclassification from all features.Comparing with using the ThreshCanny features or Graph Cuts features separately,higher classification rate was obtained by using the features based on weighted meanfusion, and the LDA, QDA, mahalanobis, KNN and SVM classifiers and leave-one-outcross-validation (LOOCV) were used for classification, in ten wood species. Moreover, the best classification result was obtained by using SVM classifier and LOOCV, withthe optimal fusion coefficient, the highest correct classification rate attained95percent.Meanwhile, the worst result was obtained by using LDA classifier and LOOCV, withthe optimal fusion coefficient, the highest correct classification rate was87percent.After feature extraction, feature analysis, feature dimension reduction and featurefusion, based on the wood transverse-section micrograph and different segmentationalgorithm,5classifiers were applied for wood classification. The experiment resultsshowed that the wood classification method was feasible and effective based on woodSEM Micrographs, using the image segmentation algorithm and feature fusionalgorithm proposed in this study. Overall, in this study, a novel method for referenceand the relevant scientific theory basis were proposed, which contributed to theresearches on wood micro feature extraction, wood classification and wood recognition.