Strip Steel Surface Defect Detection Based On A Single Image And Statistical Characteristic Of Defect-Free Images

The defects on strip steel affect the appearance of products and reduce the properties,such as corrosion resistance,fatigue strength and so on.The computer vision-based quality inspection technology on strip steel surfaces has played an important role in improving the quality of products and the capability of intelligent manufacturing.However,the existing surface inspection technologies still have many problems in the scale of training data,the accuracy of defect detection,the performance of anti-interference,and so on.In an actual production line,it is difficult to collect a set of defective samples in which the types of defects are comprehensive and the number distribution is balanced.Therefore,by applying the statistical features extracted from the test image itself or the defect-free images,three detection methods for strip steel surface defects are proposed.The detailed research contents and results are summarized as follows:(1)A Haar-Weibull-Variance(HWV)model is proposed to describe the texture features of local patches in test image and the defects on strip steel surfaces can be detected by applying the features extracted from test image itself.Firstly,an anisotropic diffusion model is utilized to preprocess the test image.Secondly,the HWV model is established to characterize the texture features of each patch in the test image.Thirdly,several patches identified as defect-free patches are extracted according to the formation of parameter distribution in HWV space,by which the HWV model parameters will be optimized further.Finally,the adaptive threshold is determined to locate defects.The proposed method which can accurately detect different types of defects and achieve an average detection rate of96.2%,which outperforms the previous defect detection methods.(2)A guidance template-based detection method for the defects on strip steel surfaces is presented,which assumes that the intensity of strip steel defect-free images obeys Gaussian distribution.For each test image,a unique guidance template is generated based on the statistical characteristic of defect-free images and intensity distribution of test image.Then,by subtraction operation and adaptive threshold determination,the defects can be located.The proposed method is a pixel-level detection method.It has obvious superiority in computational efficiency,since there exist only simple sorting operation,template generation,and subtraction operation during the detection procedure.Moreover,it can be extended to other textured surfaces by adjusting the prior distribution of defect-free images.(3)A fuzzy measures-based detection method for the defects on strip steel surfaces is proposed.It assumes that the background intensity of test image obeys Gaussian distribution.By applying the statistical intensity range of defect-free images,a Gaussian curve corresponding to background can be fitted from the test image histogram.Then,amembership function and an index function of fuzziness are defined to estimate the extent to which each pixel belongs to defect.Combining pixel connectivity,the defects can be located.The proposed method is applicable to diverse types of defects,especially the subtle defects and the defects in low contrast images.By adjusting the background intensity distribution of test image,it can be extended to other textured surfaces.