Research On Online Inspection System Of Automobile Glue Based On Machine Vision

With the proposed action agenda of "Made in China 2025",the automotive manufacturing industry is developing toward digitization,networking and intelligence.The body glue is an important part of the automobile production line.The current glue detection method cannot realize the automatic positioning of the glue,and the detection efficiency and accuracy are less.In this article,I propose an online inspection system for glue based on machine vision.This system aims both to achieve the real-time detection of defects such as broken glue strips,too wide and too narrow cross-section diameters in the production process and to realize glue position faster.At the very first,I carry out the overall design of the glue system according to the detection needs and difficulties of glue in the production process.Then make a visual sensor to collect the glue image in real time,use the neural network to realize the automatic positioning of the glue strip,and extract the region of interest of the glue strip,judge the continuity and section width of the glue strip.Determine whether to trigger the alarm signal according to the detection results,then judge whether there is reflective interference leading to misjudgment,and finally save the detection results in the database,It is convenient for users to trace and query the test results in the later stage.In order to solve the problem of large detection error caused by glue strip offset in template matching or fixed Ro I method,I use RPN network structure in deep learning network to extraction of the region of interest of glue strip,and DIo U algorithm to replace the original loss function for regression correction of the prediction frame in the training stage.In the detection stage,the NMS suppression algorithm is improved,so that it can select a more appropriate prediction frame according to the detection requirements of this paper,and realize the positioning accuracy with an average Io U of99%.The large area of adhesive strip reflection in the image will lead to inaccuracy and misjudgment in adhesive strip segmentation.The improved YOLO network is used to detect the reflective characteristics,strengthen the detection ability of reflective small targets,and improve the distortion caused by the zoom of the image by the YOLO network.Finally,75 reflective interferences in 77 reflective samples can be accurately detected to achieve 97% detection accuracy.Design the calculation algorithm of the average width of the adhesive strip section,use the BFS algorithm combined with the target confidence points to calculate the adhesive strip area in the image,refine the adhesive strip,extract the skeleton and obtain the length information;Finally,the actual width of the adhesive strip is obtained by using the mapping relationship after calibration.The average absolute error of the verification algorithm is 0.23 mm and the standard deviation is 0.25 mm,the maximum error is 0.39 mm,which meet the requirements of 1mm accuracy detection.The experimental results show that the detection efficiency and accuracy of the detection system meet the needs of field production.