| Edge, which is the mutation part of gray in the image, is a basic feature of the image. Therefore edge extraction is an important procession in image analysis, widely used in image segmentation, pattern recognition, reverse engineering, and so on. Among the verge of traditional extraction methods, the maximum gradient or the second derivative of the zero-crossing points were detected, wavelet multi-scale edge detection, and surface fitting methods were used. Neural networks edge detection as a new development of the edge extraction with advantages of parallel computing, non-linear mapping ability and adaptive capacity, and thus have attracted more attention.Industrial CT (Computerized Tomography) technology is an advanced NDT technology which scans through the work-piece, then reconstructs image sequence of work-piece slice; thereby detect defects and the internal structure of the work-piece, applied to the aerospace, aviation, railway transport, machinery manufacturing, and other fields. However, due to factors such as cross-talk noise and other artifacts, continuous edges were hardly extracted from traditional edge extraction method. At the same time, industrial CT images edges compared with natural ones have more geometry patterns, composed of lines, shapes, such as circular. In this paper, these industrial CT image characteristics were taken into account, we discussed several edge detection algorithm based on neural networks, such as BP network, CP networks, cellular neural network (CNN). The experimental results of these types of algorithms were given.In the edge extraction by BP network, one image which has the known edge serves as a training image, then training the BP network to obtain weight matrix that is used to detect the edge of the other images. It is necessary to consider how to select training images to obtain good generalization. At the same time, BP algorithm easily lead to confusion in the minimum, thereby increasing the network time expenses in training, and may never get to convergence point. On account of these problems, learning samples were constructed to detect the edges of binary images. Experiment shows the learning samples avoid convergence problems from excessive samples. For gray-scale images, first divide the images into 8 binary planes with different gray level, and then naturally the images surface up through synthesizing the edge of each binary plane. Applying the method to industrial CT images' edge detection, fine continuous edge can be acquired, the virtues of which lies in its strong resistance against noise pollution and can be widely adopted in practice.CNN used in the extraction Edge, only a small number of non-continuous edges can be got using a single CNN, which trouble the follow-up image processing. It is proposed to use two Cellular Neural Networks to segment Industrial Computerized Tomography Images. One applies image segmentation roughly to obtain thresholding images. The other segment images exactly get edges. Experimental results demonstrate the efficiency of the methods presented in the article able to get successive and fine edges.Finally, CNN edge detection was studied and improved, to extract the internal and external surfaces (called edge surface) of the scanned work-piece from the industrial CT volume data, even the surface was included in a slice. After getting the corresponding slice sequence along three mutually perpendicular directions, the edge extraction of a slice was realized by two sets of cellular neural network. Then the slice edge data were restructured to get the edge volume data along one direction, which would be integrated in three directions to gain the final edge surface. Because the gray change of a point along three directions were taken into account, this algorithm not only can extract successive and fine slice edge in a slice, but also can get more complete edge surface than dealing with only one direction. Computer experiment results validated the validity of the algorithm. |
PDF Full Text Request