Research On Recognition Method Of Wood Surface Defects Based On Gabor Transformation

Wood is an important and indispensable resource in the national economy, but in the process of being there will be many kinds of defects inevitably, and it has a very important significance for the rational utilization of wood and improving the automation degree of wood products processing industry to detect these defects nondestructively and identify them automatically. Therefore, three common types of timber defects of worm sting, live knot, and dead knot are selected as objects in this article, and produce an experimental samples library containing three types of defect samples image 50 pieces respectively, then study the identification of them by using digital image processing and pattern recognition theory.Gabor transform is a hotspot research in the image processing field, and it is of good biological visual characteristics, also, it is able to obtain optimal localization in spatial domain and frequency domain simultaneously, offering a new tool to image processing. Therefore, the method of Gabor transform is adopted in this paper to extract characteristics of wood surface defects, and then begin the segmentation and identification on them.To realize the identification of defects, the first step is to segment defect images to get defects targets. Each defect image is multi-channel transformed by annular Gabor filter firstly, then feature parameters of each pixel are defined in energy sense, forming a six-dimension eigenvectors, and the fuzzy c-mean clustering algorithm is combined to finish the clustering of each pixel. For the five categories by clustering segmentation, the categories of defect targets are determined based on their morphological characteristics and defect targets are post-processed by mathematical morphology operation. Eventually, the complete defect targets are got.According to the grey system theory, an image segmentation quality evaluation model based on grey relational analysis is established, and it is adoptd to carry on the appraisal to segmentation result of wood surface defect images, obtaining the average segmentation accuracy of the three types of defects images of worm sting, live knot, and dead knot are 95.81%,96.52%,94.58% respectively, and the division effect is satisfactory.In recognition of wood surface defects, both the features of frequency and shape are considered, and 12 Gabor parameters and 2 shape parameters of defect targets are withdrawn, and the two constitute a 14-dimension classification characteristic vector of timber images. The feature selection method based on genetic algorithm and the recognition rate of 1-NN classifier is used to optimize the classification characteristic vectors. In order to compare the recognition effect under different parameter systems, three characteristic systems are chosen as input characteristic vectors of classifiers which contain single 12 Gabor features, the combination of 12 Gabor parameters and 2 shape parameters, the fusion features after the feature selection to the 14 combinational features. BP neural network classifier is used for the recognition, and the respective recognition accuracy of the three characteristic parameter systems is 76%,81.33%, 89.33%. Finally, the 4 fusion features after the feature selection are determined as the input of the classifier. The corresponding classifier is BP neural network classifier with three layers and 5 hidden layer nodes, and its training method is Bays regularization method (trainbr).The research shows that it is feasible to segment and identify wood surface defect images automatically according to their Gabor frequency characteristics. Also, the research will provide certain reference for the later thorough research on nondestructive testing technology of wood surface defects.