Research On Key Technology Of Body In White Surface Defect Detection

As one of the important processes in automobile production,the inspection of the bodyin-white surface defects directly affects the production efficiency,quality and cost of automobiles.Due to the large surface area,complex surface structure,poor imaging characteristics of materials,and various types and shapes of defects,how to realize automatic inspection still faces many challenges.In this paper,based on deep learning technology,the key technologies of automatic defect visual inspection for body-in-white surface are studied.The main work is as follows:(1)The design and optimization of image acquisition system for small curvature area of body-in-white surface.Based on the structural characteristics of the body-in-white surface,the body-in-white surface is divided into some different regions.Through a series of illumination experiments,an image acquisition system is designed to meet the requirements of detection accuracy.(2)A supervised defect detection method with multi-scale feature analysis capability is studied to solve the problems of large size variation and low contrast of defects on body-inwhite surface,which has the ability of detection and localization.Firstly,the multi-scale generation module is used to fuse the features of different scales of the image,then the segmentation task is used to guide the feature extraction module to extract effective features,and finally the classification sub-network is used to output the anomaly score of the image.With the introduction of multi-scale generation module,the detection ability of the method for small targets and low contrast defects is enhanced.Furthermore,by introducing the hybrid supervised training method,the problem of pixel-level labeling with supervised method is effectively alleviated.(3)In order to solve the problem that the supervised defect inspection method needs to make a large number of defect training samples,an unsupervised defect detection method based on wavelet reconstruction is studied.As the surface of body-in-white is random texture,and current mainstream unsupervised defect detection methods based on background reconstruction have poor reconstruction ability for this kind of image,this paper realizes a network model with stronger reconstruction ability by fusing dual-tree complex wavelet transform and deep neural network,thus suppressing false alarming caused by reconstruction residual effectively.In addition,in view of the problem that various classical image similarity comparison algorithms can’t cope with all kinds of contrast defects at the same time,a segmentation network is adopted to replace the classical image similarity comparison algorithm,and a stronger non-linear anomaly mapping ability is realized.(4)Off-line and on-line experimental platform construction and algorithm performance verification.Firstly,according to the design of the image acquisition system,an off-line experimental platform is set up,and the off-line performance of the defect detection algorithm in this paper is verified by taking the metal samples with similar imaging characteristics as the test objects.Secondly,the dynamic detection environment in real scene is simulated,and the online detection performance of the defect detection system is verified.Above all,this paper has completed the design of automatic defect detection system for the main areas of body-in-white surface,and carried out simulation experiments by setting up an experimental platform,which verified the feasibility and detection performance of the defect detection algorithm.This work provides a feasible solution for the application of automatic defect detection system on the surface of body-in-white.