3D Detection Of Surface Defects Of Metal Strips And Plates Based On Photometric Stereo

The surface quality of metal plates,such as steel plates,aluminum strips,etc.,is an important indicator reflecting the quality of products.Surface inspection systems based on machine vision technology enable manufacturers to detect,track and grade surface defects on products,which improves quality and productivity,and also provides critical data for quality assessment and traceability.At present,on-line surface inspection systems generally adopt two-dimensional detection methods based on grayscale images.Although these systems have good defect recognition effects on steel strips,there are two problems.Firstly,there are a large number of non-defective interferences,such as scales and water stains,on the surface of medium and heavy plates,continuous casting slabs,etc..These interferences are always misidentified as defects as they are similar to real defects in two-dimensional images,such as cracks,roll marks,etc..Secondly,metal strips with high surface quality requirements,such as automotive sheets,aluminum strips,etc.,are not allowed to have small defects on the surface,which are difficult to be detected in 2D images.High misidentification rate of defects on steel products with complex surface and low detection rate of small defects on high quality strips have become a difficult issue in the field of surface inspection.Traditional 2D detection methods are almost impossible to solve these problems.3D detection methods can acquire 3D surface data,which are used to detect and classify defects to improve the detection rate and recognition accuracy of defects.As a method of 3D surface reconstruction,photometric stereo has advantages of high resolution,fast speed and simple system design.In this dissertation,3D surface reconstruction based on photometric stereo was deeply studied,and 3D detection methods of metal strips and plates based on photometric stereo was proposed and applied to surface defect detection of medium and heavy plates,aluminum strips,etc..The main results and innovations are as follows:(1)A fast 3D reconstruction algorithm based on wavelet reconstruction was proposed.The improved multi-scale Haar wavelet reconstruction operation was applied to reconstruct 3D shape from gradient fields,to speed up the algorithm of 3D reconstruction.The time complexity of the algorithm is linearly related to the pixel resolution,which is applicable to 3D reconstruction with high resolution images.3D reconstruction time of an image with 2048X512 pixels is reduced to 0.3 s,which meets the requirement of online detection.Moreover,the reconstruction algorithm with incomplete gradient was studied,and a 3D feature extraction method based on directional projection was proposed.(2)An identification method of defects and non-def'ects based on 2D and 3D feature fusion was proposed,and applied to surface inspection of medium and heavy plates.The surface gradient and depth maps of steel plates were obtained with 3D measurement based on photometric stereo.Combined with grayscale images,real defects and non-defects were correctly identified.The algorithm was tested with the imbalanced sample set,which is similar to the distribution in practical application.The results showed that the fusion algorithm reduced the adverse effects of the imbalanced distribution.Precisions of cracks,indentations,shells,etc.were increased 4%to 13%,and the overall accuracy was increased 4.9%.Water,oil stains and other non-defects without 3D deformation were scarcely misidentified.Moreover,the algorithm architecture was optimized,and the gradients were applied to extract regions of interest,which speeded up the algorithm about 57%.(3)A photometric stereo method was proposed to detect micro-defects on aluminum strips.By dual directional illumination,the effective illuminate range of micro-scratches was covered,and the contrast of micro-deformation defects was enhanced in gradient and depth maps.The resolution of defect detection reached 0.2 mm,and the detection rate was increased by 10.6%.Particularly,detection rates of point defects,horizontal and slanting scratches and micro-deformation defects are significantly improved.(4)A local monotonicity function of relative grayscale was defined to solve the directional gradients of 3D surface,and it had a wide application range compared with the traditional algorithm based on the ideal diffuse reflection assumption.Compared with the traditional algorithm based on hypothesis of ideal diffuse model,the proposed method was suitable for any reflective materials.The symmetry of the imaging scheme was considered to simplify the reflection model and calibration,and optimized the angle of incidence of photometric stereo scanning system.Smaller incident angle(±17°)was advantageous to improve the detection rate of small defects on aluminum strips,while relatively larger incident angle(±25°)enhanced imaging quality by avoiding over-exposure of bare plate surface.The above methods have been applied to production lines of medium and heavy plates,aluminum strips,etc.,and achieved good application results.