Research On Identification Methods Of Real Wood Floor Defects Based On Morphology And Clustering Algorithm

The surface defect situation of real wood floor directly influences the level of the products. It has great significance for classification of real wood floor and improving the production automation degree to realize nondestructive testing and automatic recognition for defects of different types. According to the characteristics of the surface defect of real wood floor, we therefore use digital image processing techniques and pattern recognition technology to put forward identification methods of real wood floor defects based on Morphological and clustering algorithm.In this paper the surface defect that we study was sorted to four types:dead knot, live knot, wormhole and craze. The main contents include:image segmentation for surface defect of real wood floor, feature extraction, feature selection, characteristic dimension reduction, classifier design.In the image segmentation, we put forward two methods about image segmentation in this paper. The first one is called image segmentation method based on the fusion of gradient operator and Otsu global threshold. This method not only resolve the problem that there is some deterrent when we use the image segmentation method based on Otsu global threshold to detect dead knot, but also make up the shortage that there is some omissions when we use the image segmentation method based on gradient operator. The method can achieve the image segmentation for defect when the situation of the surface of real wood floor is simple. Another method is called image segmentation method based on Morphological reconstruction. It is the focus of this paper. The method extracted the growth area image and seeds image according to the original gray image. The method removed interference through the seeds point optimization process. Then it completed the regional growth quickly through the morphological reconstruction and some necessary morphological operation. The method can overcome the shortcomings of the first method. It can realize the image segmentation for defect in most situations.In the feature extraction, we extracted three kinds of feature about image of defect: geometric and regional feature, gray texture feature and moment invariants feature. There are total twenty-three features. Then we compared their variance to select the features which we extracted. The basic principle of the comparing is that the variance of the same features of defect which classified the same kind is as small as possible and the variance of the same features of defect which classified different kinds is as large as possible. We chose sixteen features from twenty-three features as the final features in order to improve the overall speed. Finally we used the principal component analysis to reduce the dimensions of the features from sixteen to eight in order to improve the speed of the classifier. The features whose dimensions are eight are the input of the classifier.In the classifier design, we used twenty vectors of the features from the four kinds of defect to train Self-Organizing feature map which its structure is hexagonal and its size of the competition layers is10*10. We realized the classification for surface defect of real wood floor according to the distribution of winning neuron. The result of the classification is good. The recognition rate of dead knot is88.57%. The recognition rate of live knot is85.17%. The recognition rate of wormhole is93.15%. The recognition rate of craze is89.17%.