Research On Visual Identification Method For Strip Surface Defects

As one of the important products of the steel industry,strip steel has played an important role in defense equipment,automobile manufacturing,aerospace and other fields.Its surface quality seriously affects the performance of the final product,and therefore efforts to improve the quality of the strip surface to improve the performance of the final product are of great significance.Based on machine vision,this paper studies the surface defect identification technology of strip steel,and puts forward the overall scheme of the surface defect detection method of strip steel.According to the performance index,the hardware selection and parameter calculation of the image acquisition module are completed.The principle and program design of image filtering algorithm,image enhancement algorithm,feature extraction algorithm,feature selection algorithm and defect recognition algorithm are mainly studied in this paper.The main research contents are as follows:(1)Under the image preprocessing stage,focused on the common image enhancement and determined the best optimal scheme of image enhancement through comparative experiments,laid the important foundation for the subsequent defect recognition.For site dust disturbance noise is introduced,considering the filtering effect of denoising methods and combining experimental analysis,the homomorphic filtering was finally chose for image noise reduction processing.(2)A number of texture features are extracted from gray image,gradient map and Gabor transform image.As a result,a defect image is characterized by 63 features.Aiming at the problem of high dimension of feature space,the method of feature selection is adopted to achieve the goal of reducing dimension.Based on the principle of principal component analysis,ReliefF algorithm and autoencoder,the corresponding program of these algorithms is designed.These three algorithms are used to select the eigenvectors of the original 63 dimension.Building appropriate recognition model can effectively improve the recognition result,this article designed recognition model based on the BP neural network,support vector machine(SVM),extreme learning machine(ELM)and stack sparse autoencoder algorithm.(3)The comparative experiment analysis was carried out in four eigenvectors respectively combined with four identification models.The system performance of theidentification accuracy,recognition time and system stability are considered.The result show that ReliefF algorithm combined with support vector machine can not only reach98% accuracy but also has the shortest time,what is more,it has the most stable system operation.