Machine Vision Based Automatic Surface Defects Inspection For Cold Rolled Steel Strips

The quality control is a critical issue and becoming more and more important in the steel strip production. As an advanced technique in the quality monitor, the strip surface inspection system (SIS) has drawn much more attention from iron and steel enterprise. The machine vision inspecting technique, which has the superiorities of rapidness, high accuracy and reliability, is the focus of current research and application of the strip SIS. However, the technology is still in the development stage at present. Currently, the SIS research mainly focuses on reducing the system complexity and enhancing the system stability to shorten the developing cycle and save the expense, optimizing the strip defect inspection and recognition technology to improve the detecting rate of and the recognition accuracy of the system. The thesis will conduct in-depth research on SIS designing for cold rolled steel strip with key technology of image processing and image recognition.Based on the analysis of function and capability of SIS from strip mill and the summary of the current research status, a practical SIS scheme using machine vision technology is proposed. A series of main modules are designed on the system layer, as well as the system hardware and data processing in software. A parallel scheme is employed to realize acquiring real-time image by several high-speed line-scans CCD(Charge Coupled Device), and each of them connects to an image-processing computer to complete the inspection separately. The software uses hierarchical processing logic, which combine with the in-time processing and on-time processing, to meet the requirements of precision and speed. Two critical tasks involved in the defect detection of strip surface are the strip surface image processing and defect recognition.The image processing of the strip surface defects realized the task of defect detecting and image segmentation. For unstable image quality and bright in middle of image and dark on the left and right side while acquiring image by line-scan CCD, a gray compensation method is utilized, which not only can improve image quality and normalize image gray distribution, but also realize the independent development and test of the following processing. According to the character of distribution randomness and shape complexity of strip surface defects, image segmentation based on edge or area usually cannot generate accepted results. According to the local area gray distribution information, an image segmentation method has been setup. On the bais of complex shape defects and stable background in strip image, with the aid of diving image background to define the rest as defect region in the image. Tests on some typical strip defect image show that this method is feasible and efficient.Image features extraction and selection are the premise of the defects classifier built-up. A number of image features extracted from different image status, different regions and different kinds of feature description operators, which can describe the defects in feature space precisely. Feature selection should match the building of classifier. According to this theory, a filter selection scheme is present, which is the union of ReliefF algorithm and correlation analysis method. The filter feature selection can remove the irrelevant and redundant characteristics to reduce the dimensions of features, which can simplify the complexity of the design and maintenance of classifiers.The defect recognition of strip is a complicate problem with multiple types and features, and using one classifying technology to achieve a one-step classifier with fine performance will be extremely difficult. This thesis summarized the technology of integrated classifiers and graduation classifiers, and then proposed a multi-classifiers combined strip defects recognition method, which combined SLIQ decision tree classifier by Boosting algorithm. Companying with the similar type of classifier integration technique, Boosting can increase the precision of the weak classifier effectively through the adaptive weight update technology and voting with weight technology. Tests on real strip defects show that the recognition accuracy reaches above 92%, and improves the accuracy of single classifier up to 10%.The achievements in scientific and experimental research for strip SIS, such as system design, methods of defect inspection and recognition, has already obtained confirmation in the engineering project, and invested into the industrial application, which have inspiration affection on the other strip surface inspection by machine vision.