Application Of Multi-scale Geometric Transformation In Metal Surface Defect Recognition

The surface defect of metal strips is an important factor affecting product quality.A large part of the quality problems of plate and strip products is caused by surface defects.The surface inspection system based on machine vision can detect the surface defects on-line and feedback the information in time,which has become the main inspection means of modern iron and steel enterprises.At present,the main problem of surface inspection system based on machine vision is that the defect recognition rate is low.How to further improve the defect recognition rate under the condition of meeting the inspection speed is an important research content.The quality of feature extraction directly affects the application effect of surface inspection system,so seeking an effective feature extraction method is the key to improve the defect recognition rate.Metal surface defects have different information in different directions and scales,and the information of profile and geometry of the defect are mainly found on coarse scales,while the details of local edges and gray-level brake points on fine scale,so extracting multi-scale and multi-directional information of surface defects is important to improve the defect recognition rate.Compared with the wavelet transform,the multi-scale geometric transform method has more directional selectivity and the basis function satisfies the scale relationship of anisotropic.It can approach the signal singularity with fewer coefficients and better approximation order when describing the high-dimensional signals.Therefore,according to the characteristics of surface defect inspection of pickled steel strips,hot-rolled steel plates,aluminum plates and high temperature continuous casting slabs,corresponding inspection algorithms based on multi-scale geometric transformation are developed in this paper.The main research contents and innovative achievements of this paper are as follows.(1)Aiming at the characteristics of high speed of production line,simple image background and few defect types of pickled steel strips,a feature extraction method based on Contourlet transform and KSR(Kernel Spectral Regression)is proposed.The Contourlet method were used to decompose image into multiple subbands,and then the redundant features were removed by the nonlinear dimensionality reduction algorithm KSR.the obtained low-dimensional features are used for classification and recognition.Compared with several commonly used feature extraction algorithms,the recognition rate of pickled steel strips surface defects by Contourlet-KSR method is 96.76%,which is 3.05%,15.66%and 5.14%higher than wavelet method,local binary mode method and amplitude spectrum method,respectively.(2)Aiming at the problem of image information loss caused by subsampling operation of Contourlet transform method,an improved contourlet transform(ICT)method is proposed.Based on it,an ICT-KSR feature extraction method is proposed and applied to the surface defect recogniton of hot-rolled steel plates.The method of Contourlet transform is improved as follows:non-subsampled pyramid(NSP)is used in multiscale decomposition to eliminate the frequency aliasing and information loss caused by scale decomposition,and then the image is input into the dual channel directional filter banks(DFB)for directional decomposition.The recognition rate of ICT-KSR method for each kind of defects of the hot-rolled steel plates is over 97.62%,and the overall recognition rate is 99.21%.(3)Aiming at the characteristics of low contrast,small defects and many interference factors of aluminum plates,a feature extraction method based on NS ST(non-subsampled shearlet transform)is proposed.The sample images of aluminum plates are processed by contrast stretching,then the horizontal cone and vertical cone subband images of NSST are processed.The statistical features of mean and variance are extracted from each subband coefficients to form a NSST feature matrix,and the nonlinear dimensionality reduction algorithm is used to remove the extracted redundant features,finally,the low-dimensional feature matrix and labels data are input into the classifier for defect recognition.Using the features extracted based on NSST method,the defect recognition rate reached 97.92%.(4)Aiming at the characteristics of slow running speed of production line of continuous casting slabs,complex image background and many interference factors,a feature fusion method of DNST(discrete non-separable shearlet transform)and GLCM(Gray level co-occurrence matrix)is proposed.First,the continuous casting slabs samples are decomposed by DNST,then extract the statistical features of each subband image.Sencond,the gray co-occurrence matrix is calculated for the continuous casting slabs samples and five texture parameters such as energy,entropy and moment of inertia are extracted,then the mean and variance statistics of the five texture parameters are calculated respectively.Third,the statistical features of both DNST and GLCM are connected to form a feature fusion matrix.Finally,nonlinear dimensionality reduction algorithm is used to remove the rudundant between features.Using this feature fusion method,the overall recognition rate of continuous casting slabs defects reached 96.37%,of which the crack recognition rate reached 95.50%.(5)Taking advantage of the phase information of complex shearlet(CST),a defect detection algorithm based on complex shearlet transform is proposed and tested on slag inclusion defect of pickled steel strips and scratch defects of aluminum plates.The algorithm can accurately locate slag inclusion defect of pickled steel strips and scratch defects of aluminum plates.According to the characteristics of continuous casting slabs surface defect recognition,a feature fusion method based on CST is proposed,which makes the recognition rate of continuous casting slab defect up to 95.97%.