Research On The Detection Method Of Highly Reflective Arc Surface Defect

Metal products with high reflective arc surfaces are widely used.It is difficult to detect surface defects based on vision,since the arc have strong specular reflections and viewing angle occlusion.In this paper,we designed a pipeline surface defect detection system for tube metal products with high reflective arc surface.We also propose a variety of detection methods based on visual analysis and comparative analysis.The specific research contents of this paper are as follows:1.First of all,we analyse the influence of high reflective surface defect detection.To solve the negative impacts of the defect detection from our analysis,this paper summaries the advantage and disadvantage of high reflective surface defect detection system,which incorporated the fixed multiple points of view method.After analysis,we decide on designing a highly reflective surface defect detection system based on work-piece rotation procedure.Then,we analyzed the advantage and disadvantage of using different types of camera,light source,and illumination intensity on the defect characteristics of high reflective arc surface.From which we obtained the optimal illumination intensity.Finally,to complete the design of defect detection system,we discussed and found the most appropriate angle among light,camera,and work-piece.2.First,multiple denoising methods for the high reflective tube surface with strong reflective characteristics are compared,and finally,of which the bilateral filtering is chosen as the image denoising processing method.Second,for image segmentation,a fixed threshold method is chosen to obtain defect features through comparison experiment of multiple image segmentation methods.Third,after extracting the feature contour,two defect detection methods are proposed.One is based on the minimum enclosing rectangular template matching,and the other one is based on the generated template matching.Experiment results show that the previous method is with rapid reaction due to its simple structure,and the later one has a huge advantage including higher accuracy and detect location function.3.In Chapter IV,we firstly introduced the basic theoretical knowledge of convolutional neural network.Then,we analyses the various activation functions and optimization algorithms,and Re LU function Adam optimization algorithm are chosen as the network model activation function and the network model optimizer,respectively.Multiple CNN structures like VGG-16,Res Net-18 and Densenet-121 are applied and validated by experiment.With the consideration of application for embedded system,an improved VGG,called as S-VGG,is proposed.Experiment results show that S-VGG is with higher performance compared to VGG-16,Res Net-18,and Densenet-121.Moreover,S-VGG has an obvious advantage of application for embedded system due to its simple structure and less parameters.