Research Of Key Technologies For Steel Plate Surface Defects Detection

There are many kinds of surface defects during the production of steel plate. The main factors causing these defects are devices, materials, technologies, casting and rolling. So, for steel and iron enterprises, technologies of surface defects detection are used to monitor the quality of steel surface. Detection technologies have developed for over ten years and achieved some achievements in steel and iron enterprises and research organizations which are at home and abroad. However, as the years go by, the requirement of market is improved. And specially, with the development of IT, internet and cloud services, steel and iron enterprises will be inevitablly impacted. In order to meet the above changes, home and abroad steel enterprises started to detect and control the surface quality of steel plate on the wohle production line. On one hand, in order to meet the requirements of high quality, steel and iron enterprises must improve the criterion of quality monitoring for steel plate surface. Meanwhile, in order to meet the requirements of high-speed production, steel and iron enterprises have to improve the efficiency of the production line. On the other hand, surface defects detection system should not be equiped for an independent production line, but for multi-strip production lines, which can realize the reuse of resources and optimized equipment. However, the current research achievements at home and abroad enterprises can not satisfy those high requriments. So, under the requirements of high efficiency, high accuracy, multi-strip production lines and IT servise, it is very necessary to study the detection technologies for steel plate surface defects.Technologies of steel surface defects detection are studied in this article and the key contents and achievements are the following:(1) Under the requirements of market and the development of enterprises, new indicators and requirements of detection system for multi-strip production lines are firstly determined based on the analysis of independent steel plate surface defects detection system. Then, clustered type of steel plate surface defects detection system is proposed. Its structure, function, equipment and operation are analyzed and designed from acquisition, transmission, clusters and memory respectively in this article. Moreover, its detection flow is analyzed and designed for region of background detection for suspicious defect images, region of defect detection for defect images, defect detection, splitting and allocating operation, and defect information processing.(2) A novel region of interest (ROI) detection method is proposed to match the clustered type of detection system. The algorithm of ROI detection is divided into two steps. The first step which is called region of background (ROB) detection is to build standard background images database and local projection values database. The algorithm which combines the parallel projection and substraction methods can satisfy the on-line requirements and reduce the miss detection rate. The second step which is called region of defect (ROD) detection is to extract projection values in four directions from local windows of filtered images. And five statistical features are obtained from all the projection values. Then extreme learning machine is used to realize the classification between ROB and ROD, which can not only ensure the miss detection rate but also reduce the false detection rate. Experiments show that the ROI detection method has the characteristics of low miss detection rate, low false detection rate and high efficiency.(3) Related research of position detection for defect region is also done. A novel filtering method is proposed in this part. Five-median-binary code (FMBC) of local edge model is used to filter salt-and-pepper noise. Then local enhanced bilateral filter with FMBC and a new type of exponential weighting function is used to remove Gaussian noise. Experiments show that this novel filtering algorithm can not only filter mixed noises of defect images but also reserve more edge details. The approach to the segmentation of ROD including differential operator, canny operator and watershed algorithm is studied in this part. And improvements are done for differential and canny operator. Experiments show that these methods are fit for ROD segmentation. Moreover, the segmentation results are used to design position information.(4) Based on the center of gravity of ROD, invariant resampling is relized by using the concentric square sampling template with rolling. Based on resampling, three extraction methods of statistical invariant features are proposed for edge of defect and ROD. The first one is the statistical feature analysis and extraction of the edge distance for the normalized distances which are from the points of edge to the center points of gravity. The second one is the statistical feature analysis and extraction of the gradient orientations obtained from two adjacent points of edge. The last one is the statistical feature analysis and extraction from smoothing local binary pattern values with eight weighted grays. Experiments show that those three types of features can describe the ROD well, which provides more features information for the following defect classification as much as possible.(5) Based on twin support vector machine, Multi-density TWSVM (MDTWSVM) model is proposed. This classifier uses density information to estimate the importance of sample, which realizes the increasing and decreasing of samples for unbalanced datasets. Density information is used to improve the object function of TWSVM. Successive Overrelaxation algorithm is used to resolve MDTWSVM model fast. Binary tree and MDTWSVM are combined together to realize multi-classification of steel plate surface defects. In addition, amplification factor and weighted information of pruned samples are used to improve least square TWSVM and obtain new classification model, which can increase the classification efficiency without decreasing the accuracy and restrain the impact of noise samples. Experiments show that those two classification models can realize steel plate surface defect multi-classification with high speed and accuracy.