Research On Surface Defect Detection For Stell Strips Based On Machine Vision

Quality control in steel production has gained more and more attention .Plate surface defects that affect the quality is the important factors for strip surface .Defect detection for improving the quality of plate plays a very important role. Detection of the traditional manual method can not guarantee higher detection level and detection level of consistency because of examiners are vulnerable to their own skill level, emotional, fatigue, as well as the impact of external factors. Artificial testing has been unable to meet demand,the research and development of automatically strip surface defect detection system has become the consensus of the majority of iron and steel enterprises. Since the 1990s electronic machine vision technology and the rapid development of technology-based machine vision technology on the surface of strip surface has gradually become the mainstream of nondestructive testing technology, in the actual system has some very good applications. In this paper, which is based on machine vision strip surface defect inspection systems, image processing software to meet the requirements of real-time detection, and the system design, which is based on artificial neural networks and genetic algorithms to identify and classify defects network to a certain extent learning ability. When the analyte material changes in the system,we can defect sample for the identification of defects and classifier training to adapt the system to changes in the related production lines.This paper studies the three-tier network of artificial neural network system for the identification plate image defects. If there are defects, the image will be saved into the buffer for the next phase of defect classification.This paper mainly on the following aspects:1,Hardware structureThe hardware framework of based on Machine Vision plate surface defect detection system majorly contains lighting equipment, CCD camera, image processing computers, servers, and LAN components. Images collected by the CCD camera spread to computer image processing by twisted pair to image processing and pattern recognition. Results of Recognition will be saved into the database server for further processing . In the future , the scene can be used to generate production reports.2,Image Processing1) Image preprocessingThe image shooting the scene from strip surface is affected by, lighting of environment, as well as factors such as the camera itself. These factors will cause some noise in image. Noise have a significant impact on the later identi -fication of texture, shape, gray feature extraction ,will bearing on the accuracy of the identification results directly. Therefore, noise filtering needs in the image for further processing.In this paper, using median filter to eliminate high-frequency noise image. Median filtering uses a sliding window with odd points, uses the gray value of the windows to replace the designated point (the center window) gray value.2) Histogram equalizationThe main purpose of histogram equalization is highlight the main content of images with the original pixel higher frequency range wide, and the lower frequency of compression of pixels. First, statistic the number of pixel of all levels of gray, and calculated distribution of the original histogram, then calculated the cumulative histogram distribution tk Formula tk = [(N-1) * tk 0.5] from the whole and that its original gray sk tk to the gray mapping. Repeat the above steps, the source of all the gray-level image to the target image of the gray-level mapping, again in accordance with the new mapping between the source image to the point pixel gray-scale conversion, to complete the straight side of the image Figure equalization.3) Image SegmentationImage segmentation divided orign image into each of the regional . At present, there is not a common image segmentation method and there was no image segmentation judge the success of the objective criteria. And the site specific defect information to determine segmentation algorithm. This paper presents a method based on the poor image segmentation method with the two point-to-point subtraction , which are the output of the image .4) Regional markersWe need to identify the location of defects in each image in the buffer and analize each region to determine which defects can be combined into a single one before treatment, Regional markers including positioning and clustering, algorithm is rather complicated. It's usually need two steps image scanning in traditional regional labeling method, and the efficiency is very bad, especially in the irregular images. This method is stimulated by the thinking of regional growth, not need progressive scan. But to mark the one-time regional connectivity, and then to mark another region, until all regional connectivity have been marked.5)Feature ExtractionTo extract defects'effective regional characteristics, as the input of identification and classification. In this system,I has extracted about 17 defect characteristics, including Gray, texture characteristics, statistical characteristics of fractal characteristics, and so on. These features is the mathematical description of surface defects, The system extracts about 17 defect characteristics, including Gray, texture characteristics, statistical characteristics of fractal characteristics, and so on. These features strip is the mathematical description of surface defects, for example, the average gray level, the average gradient than the average gradient shading domain characteristics, and so on. Mathematical description of the complete image of the less information lost in the same classification for classification under better circumstances. It must be noted that the more features does not mean that the description of the defects complete, and sometimes caused lower classification .3,Clifton, classifierRecognition of the role is based on the extraction of the characteristic parameters of plate image. The classification is based on extraction of the characteristic parameters of the defect image.In this paper,I studied Clifton and classifier based on artificial neural networks and genetic algorithms, which has a certain ability to learn ,more fault-tolerant and adaptability. The results of classification will be saved in the server's database for further statistical analysis.