Wood Defect Detection Based On Clustering Analysis Technology

In wood materials processing production, our wood defect detection technology and equipments are far behind those developed countries. Traditional detections based on physical methods have problems like great cost of testing equipments, high requirements for physical environment in actual testing while auto detection and location of wood defects with machines can reduce influences due to subjective factors like moods and fatigue etc., but speed and accuracy of identification are to be improved.Aiming at problems in traditional wood defects detection, this paper studies various abstractions of wood image characteristics, raises auto identification technology of wood defects based on unsupervised clustering which includes the following aspects:1. Recognition of wood surface defects based on color matrix features and clustering analysis technology. Abstracting color matrix features of wood surface image, it applies K-MEANS of clustering algorithm to identify wood surface defects automatically. With statistics of identification efficiency of different defects styles, it shows the validity of low level color matrix feature abstraction and K-MEANS algorithm.2. Recognition of wood defects based on GLCM and clustering analysis which obtains 14 texture features of wood image based on GLCM, calculates the coefficient matrixes between the feature values and then chooses five typical ones.Texture feature abstraction based on GLCM shows better identification accuracy after testing results comparison.3. Recognition of wood defects based on BIRCH algorithm which builds CF tree within certain threshold thus to generate initial clustering. It discusses selection of the branching factors (B, L) of threshold T and non-defect judgement in detail. Compare with K-MEANS algorithm, the results shows this method have more accurate identification.4. Recognition of wood defects based on AP algorithm which abstracts color matrix features of wood image, improves image searching method, applies multi-image scanning, automatically adjusts slide pane sizes to reduce sample sets data after feature abstraction effectively, to lower dimensions from matrix S thus to make less dimensions of distance matrix, representative matrix, selection matrix and decision matrix in AP algorithm. Testing results show that improved AP clustering methods with good identification effects which result in identification accuracy and recognition speed improvement.