Research On Visual Inspection Method Of Automobile Fuse Box Assembly Quality

With the booming development of automotive manufacturing and computer vision technology,machine vision technology is widely used in various quality inspection processes of automotive manufacturing and assembly because of its advantages of high efficiency,high precision and low cost.Based on the lack of functions of the existing image inspectors system,we designed a deep learning-based component segmentation method for fuse box images and improved the texture feature recognition method for component images to improve the efficiency and stability of the system.At the same time,a stereo vision measurement system was designed to judge whether the placement status of the component is qualified by measuring its height.The paper mainly accomplished the following work:(1)Component segmentation of fuse box images.Based on a large number of fuse box image data sets with different light intensities and different product types accumulated by several customers of the original detector system over a long time,Label Img software was used to label the location of the components and classify them into three categories: small fuses,medium-sized fuses and relays.Then Mosaic data enhancement method is used to enrich the dataset,and YOLOv5 deep learning network is designed to realize the segmentation of components.(2)Component image enhancement.Based on the research of image pre-processing methods such as image filtering,edge detection and image enhancement,an image adaptive enhancement method is proposed to cope with the relay images problems such as uneven light and darkness,and poor contrast,in which the threshold value was automatically adjusted to adapt to the component images under different illumination and reduce the impact caused by uneven illumination.(3)Improvement of texture feature detection method.Firstly,the shortcomings of ORB algorithm in relay texture feature detection are summarized and analyzed through experimental research,then a feature point filtering method based on the maximum density constraint and a four-valued feature descriptor method is proposed to improve the accuracy of feature point matching,and finally a similarity calculation method based on the feature coverage area is proposed to effectively solve the problem that similar components are difficult to distinguish.(4)Component assembly height measurement.Two images of the fuse box are collected by controlling camera move to different locations and matches the component images by matching the prediction window with template.Three feature point matching methods,called contour,character and texture,were designed to adapt to different components in the fuse box,and the height calculation of component feature points was completed according to the imaging model of stereo vision.The accuracy of the measurement system is tested through a large number of measurement experiments,and a criterion for evaluating the component assembly quality is given.