Large-Scale Metal Surface Defect Detection Based On Near-Field Photometric Stereo

The reform of the manufacturing industry not only needs to improve the manufacturing process and production efficiency,but also has higher requirements for product quality control.Surface quality is an important part of product quality.The existence of surface defects not only affects the appearance of the product,but also reduces the performance of the product.Visual inspection technology based on 2D images is widely used in surface defect detection.Compared with the human eye,it has great advantages such as high resolution,objective evaluation,high automation,and longterm fatigue-free work.However,when it comes to non-Lambertian metal workpieces,2D visual method is susceptible to high reflection problems,which leads to poor inspection results.Photometric stereo can reconstruct pixel-level surface detail texture information,and is very sensitive to depth mutations caused by surface defects.It is a three-dimensional vision technology suitable for surface defect detection.This paper proposes a large-scale non-Lambertian metal surface defect detection technology based on near-field photometric stereo.This technology can effectively reconstruct the detailed texture of nonLambertian metal surfaces,and combines texture information and original image data to quickly locate and identify small defects with a minimum size of only a dozen pixels in a large area.This paper proposes a non-Lambertian metal surface texture reconstruction method based on near-field photometric stereo.This method proposes a new near-field photometric stereo network based on the nearplane assumption and the inverse reflectance model with collocated light,and designs a specific synthetic dataset rendering process to further improve the network’s adaptability of metal surface reflection and the ability to restore texture information.The results of synthetic experiment based on rendered images show that this method can restore high-precision surface normal,and the average angle error is less than 0.9°.The results of real experiment based on the metal sample with cutting texture show that the method can reconstruct the metal surface texture with high precision.Compared with the white light interferometer,the measured root mean square value of the texture depth has a relative error of less than 15%.Combining the texture information and original image,a method for locating and identifying small defects on large-scale metal surfaces based on multi-information fusion is proposed.This method first locates the area with suspicious defects through multi-size template matching in the gradient map,and then designs a multi-information fusion defect recognition network,which combines the normal map of the area and the original image to make the final identification of the defect.Experimental results based on the surface of the smoke collecting hood with defects show that this method can effectively locate and identify defects on large-scale metal surfaces,with a detection accuracy of 98.5%.Based on the proposed defect detection method,this paper builds an automatic detection system for the surface defects of the smoke hood.The system uses a robot to drive a photometric stereo vision imaging device to sequentially detect whether there are surface defects in each area of the smoke collecting hood,and realize fully automated detection.The actual operation test on the smoke hood production line shows that the system is stable and can effectively detect small defects on the smoke hood surface and judge whether the product is qualified with a detection accuracy of 98%.