Study On Wood Defects Testing Based On BP And RBF Neural Networks

Wood nondestructive testing is a synthetic, nondestructive detecting method. It can detect the wood defects accurately without destroying the surface and structures of the wood. So it can make sure to select wood properly, use wood effectively, which will enhance the occupating coefficient and economic value of the wood and save the wood resource effectively. This thesis studies the two typical kinds of north woods:larch and birch, which belong to softwood and broad-leaved wood. The three typical wood defects:knot, rot and grub-hole have been studied in detail.The X-ray wood nondestructive testing system, hardware platform image acquisition for wood defect testing is invented. The wood nondestructive testing has been done using X-ray method. The defects can be found by detecting the X-ray difference between the rays fore and after transmitted the wood. This method can not only detect the surface defects of wood, but also the internal defects. The signals are received by the image enhancement equipment on the other end of the wood, and then are transmitted into A/D converter, which convert the X-ray analog pattern of wood into digital images and are stored into computer. This is the all process of image acquisition for wood defect.Digital image processing is applied for preprocessing of wood defects images. Gray transform enhancement, improved weighted mean filtering and median filtering are applied for the wood defects images of these three kinds of trees discussed above. The results show that the contrasts of these images enhance apparently, and there is no the widened mean filtering traditionally, the image defect details are reserved maximatily, and it is easy for the image feature extraction. The several commonly used edge detection operators:Roberts operator, Sobel operator, Prewitt operator, Log operator and Canny operator, are used for edge detection of the wood defect images. The detect result shows that Sobel operator is better and faster. In order to improve the effect of edge detection, binarization processing has been applied for the original images, the gray transform enhancement images, the weighted mean filtering images and the median filtering images of wood defects respectively, which found that the weighted mean filtering images is the best. Then the defect areas of binary images of wood defects are filled up. For extracting the clear edge of wood defects, the filled binary images are edge extracted with Sobel operator.A new method for segmenting the wood defects images by image adduct has been presented. First, invert the best binary image of wood defect, and convert its array into uint8 type, and then adduct the image obtained after median filtering with the defect image, which is turned over. The result of wood defect image plus is that defects regions take apart completely with their background, which means the image segmentation is completed. At the same time the characteristics of the defect regions, such as shade of gray, structure, do not changed.The choice of wood defects has four characteristics:distinctiveness, which means there should be an obvious difference for different objects; reliability, which means features should be similar for the same types of objects; independence, which means the characteristics used should be non-interacting, unrelated; small quantity, although big quantity can make it easier to distinguish different objects, but the amount of computation and the computation time are increased, which will decrease the identify accuracy. According to feature differences of knot, rot, grub-hole, the gray and shape characteristics of wood defects are extract mainly. First, scan the wood defects images after the edge extraction, record the coordinates of the edge of defect regions, determine the location and size of wood defects, then feature extraction is applied for binary images and segmented images of three kinds of wood defects. Four characteristic quantities have been extracted for wood defects:aspect ratio, circular degrees, the mean gray, gray variance. Combined the structural characteristics of knot, rot and grub-hole, geometric moment is introduced. By non-linear combination of geometric moment, a set of moment:Hu invariant moment, which is invariant for image translation, rotation and proportion is derived. According to the physical meaning of Hu invariant moment, the original Hu invariant moments are abstracted, its seven features change into ten features, but still invariant for image translation, convergent-divergent and rotation. Ten structural characteristics of wood defects are extracted. So the fourteen characteristic values of wood defects are extracted for these three kinds of typical defects, and are normalized. The normalized eigenvalues will serve as the input feature vectors of neural network.The BP neural network model has been designed, determined the number of layers of BP network and neuron numbers for each layer. The hidden layer and output layer select S-function as an incentive (transfer) function. Through the comparison of various algorithms of BP network, it is found that Levenberg-Marquardt learning algorithm is the optimal algorithm. Learn the BP neural network with known samples, simulate the unknown samples with this mature network, which find that the accuracy rate of this BP neural network for wood defects recognition is as high as 90%.RBF neural network model has been constructed, RBF neural network has good approximation for the data. RBF neural network is three-layer network, identified the nodes of each layer. Hidden layer transfer function belongs to Gaussian function, the output layer transfer function is a linear function, after repeated testing, radial based distribute function "spread" is determined. The RBF neural network is trained with known samples, and then simulation is applied on unknown samples. It is found that the accuracy rate of this network for wood defects recognition is as high as 92%. Set up a new BP-RBF mixed neural network model. BP neural network has a better data compression capability, RBF neural network has a good approximation to the data results, will be BP neural network and RBF neural network in series. BP-RBF mixed novel neural network part of the output of BP neural network RBF neural network as input, using the distribution of training methods, the first part of the BP neural network for training, and then to the RBF network training. As the BP-RBF mixed neural network and BP network portion of the BP network structure built in front of the same, had trained mature, do not need to carry out training, but a new neural network RBF network portion of the front of the RBF neural network with different pairs of RBF network portion of the training, re-established the distribution of constant RBF spread, to be part of the training RBF networks to achieve precision, the entire BP-RBF mixed neural network training is completed, to identify unknown samples. The sample recognition accuracy rate is 96%. Three kinds of neural networks for wood defect recognition rate comparison, BP-RBF mixed neural network to recognize the highest accuracy, but the actual output closer to the target output value.Experimental results show that application of BP and RBF neural network can successful nondestructive test for three kinds of typical defects of wood. The recognition rate of wood defect is higher above 90%. The formed BP-RBF mixed neural network model for wood defect is higher in accuracy and more effective. This method provides an important theoretical basis for automatic detection of defects in wood.