The Research For Recognition Method Of Wood Surface Defect Based On SVM Combined Manifold

For the non-destructive, rapid, accurate and economical advantages, the wood surface defect recognition based on machine vision theory and technology has a higher application value in the fields of timber production and processing. Pattern recognition and measurement of wood surface defects is a major problem of wood processing industry, and is also worth long-term studied. The accuracy,real-time and robustness of defect detection has become the hot research spot for the domestic and foreign scholars. In this paper, burrow, dead knot, slipknot are the most three common wood defects for the study, conducted intensive research on the wood surface defect image segmentation, feature extraction and recognition problem based on machine vision and manifold theory.Image segmentation is the fisrt issue of the wood defect recognition. For over-segmentation shortage of local thresholding method, this paper presents a defect detection method based on the integration of global visual saliency and local threshold segmentation method. Firstly, defect areas are coarse positioning by global visual saliency. Then, the wood surface defects are precise positioned and segmented using the local threshold method around the coarse positioning areas. Lastly, we use mathematical morphology tools processing the segmented image, which improve the accuracy of the defect extraction.In order to ensure the reliability of recognition results of wood surface efects, feature extraction is a crucial step in pattern recognition. The Tamura texture, GLCM, Local Binary Pattern texture feature of burrow, dead knot, slipknot defects were extracted and fused to form the covariance matrix manifold to reduce the feature dimension, and reduce the information redundancy between each features. The experimental results show that, the proposed covariance matrix manifold feature has good separability.This paper use manifold distance instead of traditional Euclidean distance, presents a new manifold based support vector machine classifier, using Which takes Local Binary Pattern and the first three features of Tamura texture, together with the entropy feature of GLCM as the input of classifier, and achieved better classification results.The research results show that:(1) Using machine vision theory and technology, as opposed to the traditional image-based segmentation algorithm edge based image segmentation algorithm region and segmentation algorithm based on a particular theory, the use of wood surface defect image globally significant features combined with local thresholding and mathematical after morphological processing, play better segmentation results, the ability to burrow, dead knot, slipknot defect sample images from the background well segmented.(2) By analyzing the characteristics of burrow, dead knot, slipknot three defects using image texture features describing the defect. The different types of texture defect area as a feature element integrated into the covariance matrix, constitute covariance manifold features, reducing the dimensionality of image features, but also reduces the redundancy features, based on regional characteristics covariance manifolds have good separability between classes.(3) Classical BP neural network classifier and vector space support vector machine classifier compared the proposed algorithm based on support vector machine manifolds achieved better recognition results, recognition accuracy rate of 91.67%, higher than the BP neural network classifier 82.75% recognition accuracy and vector space support vector machine classifier recognition accuracy rate of 84.17%.This paper study on identification of wood surface defects by machine vision theory and methods, the research results can promote wood nondestructive testing, improve the utilization rate of wood,will bring huge economic benefits to the wood processing industry, has important theoretical significance and broad application prospects.