Research On Surface Defect Classification Of Small Magnetic Rings Based On Machine Vision

As the largest manufacturer in the world,there are many magnetic materials manufacturing enterprises in China,but the quality detection technology of magnetic material products is relatively backward and cannot meet the requirements of precision and efficiency.With the development of machine vision technology,it is possible to detect the detection of magnetic material by machine vision method,which with the advantages of fast speed and high precision.In this context,This paper studies the surface defect detection of magnetic ring with machine vision technology.In view of the characteristics of detection of magnetic ring are small and difficult to detect,micro lens and surface array CCD imaging system are used to acquire magnetic ring image.In image pre-processing,aims at the uneven gray and noisy image obtained by array CCD,this paper studies several classic gray transform method firstly and selects the global linear transformation;Then analyses some denoising algorithms and chooses better-effect Gaussian Filter to remove the noise by valuing PSNR and NMSE.In image segmentation,while the classical algorithm of edge detection and Threshold segmentation can't achieve good results,the paper proposes the 2D Gabor based threshold segmentation method.This method filters the image by 4-scale 7-directional 2D Gabor filters and constructs the threshold according to the image gray average and variance,then reconstructs the defect area and connects the fracture ones.The experiments prove the method can not only segment the defect effectively,but also can restrain the noise.For feature extraction of the defect area,the paper extracts 34 features including geometrical features,gray features and texture feature,and then reduces dimensions with Fisher discriminant analysis combined with principal component analysis.The paper selects 23 features by calculating the Fisher ratio,and finally chooses 16 principal components which provided 95% cumulative contribution to train and test the classifier.Back-Propagation neural network and Support vector machine are well-known classification algorithms of pattern recognition field.The paper designs the neural network classifications for surface defect identification of magnetic ring first,while the accuracy is not satisfied.So this paper designs the SVM classifications and optimizes its parameters through several optimization algorithms.Experiments show that the parameters determined by the grid search method make SVM model the highest recognition rate,which is better than BP neural network model.