Research On Online Visual Inspection Technology Based On Photometric Stereo

In recent years,with the upgrading of the manufacturing industry,the degree of automation and intelligence has been rising,and the traditional testing methods have been unable to meet their needs.In particular,the process industry represented by the automotive industry is increasingly demanding accurate,robust,real-time,online next-generation detection technologies.Machine vision-based inspection technology has received more and more attention because of its non-contact,high precision,low cost,and online features.However,when visual inspection of objects with complex optical features is carried out,the detection of white-body welded studs is an example.Due to the reflective characteristics and geometrical features of the surface,the traditional visual inspection methods are facing enormous challenges.The accuracy and beat requirements cannot be met.In view of this situation,this paper focuses on the problem of welded stud visual inspection,which has complex optical characteristics.The main research contents of the thesis include:(1)Position calibration of point source based on cube target.Aiming at the shortcomings of the existing light source position calibration algorithm with low precision and complicated operation,a calibration algorithm and image processing flow based on cube target are proposed.The position of the point source is obtained by obtaining the shadow image of the cube target under the illumination of the light source.The calibration accuracy of the proposed algorithm is verified by experiments,and the required directional information of the light source is provided for the next photometric stereo algorithm.(2)Research on photometric stereo combined with deep learning.Aiming at the problem that the traditional photometric stereo algorithm is difficult to describe the optical characteristics of complex surfaces,the deep learning method is used to learn the mapping relationship between the brightness change of the object and the normal under different illumination sources.Since the surface normal of objects in the real world is difficult to acquire and deep learning requires a large number of samples for learning,this paper first proposes a BRDF expansion method based on linear combination,using the BRDF of 100 real-world materials provided by the MERL BRDF dataset and generating 100 kinds of "virtual materials".Secondly,images of 200 spheres of different materials at specific ray angles are rendered,and these images are used to train the PS-DenseNet proposed in this paper.Finally,the algorithm of this paper is tested on the DiLiGent dataset,the largest photometric stereo dataset.The results show that the method of the photometric stereo algorithm proposed in this paper can be used to realize the reconstruction of the non-Lambertian material,and is not affected by the change of the material.A large advantage is shown in comparison to the traditional photometric stereo algorithm.(3)Stud location detection based on deep learning.Aiming at the problem that traditional visual detection methods are difficult to accurately detect objects with complex optical features,a stray position detection algorithm combining photometric stereo and CNN is proposed.The stud image illuminated by eight different light sources is converted into the surface normal map of the stud by the photometric stereo algorithm as the input of the CNN,and the center point coordinates of the stud head and tail are used as the output of the CNN.Experiments were carried out on welded studs containing welded blackened,high-gloss mirrors,and complex thread shapes.The experimental results show that the proposed algorithm is better than the original image for position detection.At the same time,the proposed system takes less time than the existing measurement technology and can meet the requirements of online measurement.