Recognition Of Strip Steel SurfaceDefect Images Based On Parallel Classifiers Integration

With the development of economy, strip steel has become a kind of indispensable raw material in cars, home appliances, mechanical manufacturing, aerospace, chemical, shipbuilding and other industries, occupys an important position in national economic. In the process of production, different types of defects such as edge sawtooth, welding line, impurity, wrinkles etc, will be brought because of continuous billet, rolling device, machining technics etc. These defects not only affect the appearance of products, but also reduce the capabilities of eroding, abrade and other strength performance, which has become an important factor in affecting surface quality of cold-strip steel. To improve the strip steel surface quality, the first problem to deal is to inspect and classify these defects, then analyse the defects causes, and propose a method to eliminate these defects at last.The difficulty in strip surface defects recognition behaves in two aspects mainly:①a certain class of defects includes the defects of other types, such as wrinkles contains compositions of impurity.②the forms of the same classification defects exist big differences, such as wrinkles, impurity. This proposed higher requirements to classifiers. Aiming at the existed problems of strip surface defects detection methods, such as single classifier is hard to breakthrough in algorithm, the weakness of individual classifier and serial classifier integration recognition relys on training samples strongly, giving a recognition method of strip defect images based on parallel classifiers integration.Multiple classifier integration refers to construct a set of classifiers, and gives the classification results through (weighted) vote of the base classifiers. The purpose of multiple classifier integration is to make full use of every base classifier, then achieve higher recognition rate than any single base classifier.In this paper, using11dimensions features as inputs after dimensionalityreduction of26dimensions included statistical features of gray histogram etc, choosing BP neural network, LVQ neural network, RBF neural network and Support Vector Machine as base classifiers, to recognize edge sawtooth, welding line, impurity and wrinkles. Choosing final base classifiers based on their differences, then weighted integration using ballot and a weighted vote method. By experiments, indicate that is feasible to use multiple classifiers integration method in the strip surface identification. The recognition rate reaches above95%for120defect images.In this paper, using the research achievements of our laboratory about features extraction and dimensionalityreduction based on PCNN neural network. In this experiment, integrate classifiers using the features in the study, and the results achieved97.5%. In addition, using wrinkles and impurity which forms exist big differences in the experiments, and the result shows that the parallel integration classifier system not only can improve the recognition rate, but also less dependent on the training samples, its generalization is higher.