Research On The On-line Inspection Method For Glue-spreading Quality Based On Machine Vision

The quality of glue-spreading of the cars can affect the performance of sealing,noise,vibration,and waterproofing of the cars.Considering the security problem,the quality of glue-spreading must be strictly controlled.Now,the main detection methods for the glue-spreading are manual and off-line visual.And the latter one mostly achieved by comparing standard images or standardized trajectories of glue-spreading,which leaves the problem as most of them lack the detection for the whole width.Accurate segmentation of complex background images in dimensional detection is also a difficulty in the research.By analyzing the indicators of precision and difficulties of system,this paper provides an overall design idea of on-line inspection system for glue-spreading based on machine vision,aiming to provide an inspection system for detecting glue-spreading quality,which aims to be as real-time as possible,efficient and accurate,and doesn’t require any standard images or trajectories of glue-spreading for comparison.The gluing quality tends to be controlled within the accuracy requirements while applying.The difficulty of the segmentation for the complex background in two-dimensional inspection will be discussed.Based on the existing glue application robots,this paper presents a set of online inspection system for glue-spreading quality based on monocular vision,including the hardware platform and software algorithm.The paper works as follows:(1)Optimize the image acquisition device of the hardware platform.Based on the needs of online detection and the difficulties of image acquisition for glue-spreading,In addition the hardware is partly selected to meet the accuracy requirements.While the imaging scheme is redesigned for solving problems,such as reflection of glue and workpiece.The best imaging scheme is selected through experiments,and the software platform is connected through Ethernet to achieve information exchange,etc.,the building of the hardware platform is eventually completed.(2)Design algorithm of the image processing of the software and quality inspection and analysis of glue-spreading.Aiming the problems that can affect image quality including noise,platform’s vibration,and reflection of glue-spreading and workpiece that may occur in the process during capturing the image,an appropriate processing method is selected to process the image to provide better quality image for subsequent work.Based on the features of the detected workpiece,the design of the controlling method for glue-spreading position is completed from workpiece and demonstrated point,and then detection algorithm for the glue continuity and size are also designed.The glue-spreading quality detection algorithm is completed.(3)Combine the U-Net Semantic Segmentation to expand the application scenarios of the system.Aiming at the problem that the two-dimensional detection in the traditional segmentation method performances poor for the target segmentation from complex glue-spreading trajectories and complex scenes,the U-Net network is introduced to improve the segmentation efficiency and versatility of the system.In order to ensure that more accurate segmentation results can be obtained when the data set is small,and to reduce the workload of image collection after the system is expanded,the 2 sets of complex background images of the glue-spreading workpiece are collected for 30 pieces each,and the data is enhanced to 200 pieces,then of which150 pieces are trained.This method can achieve a good segmentation effect even on the complex background for the glue-spreading inspection,and pave the way for the subsequent detection algorithms.The research starts from the design of the imaging scheme,and then designs the detection algorithms from the position,continuity and size of glue-spreading,lastly introduces the U-Net to expand the application scenarios for the complex background in two-dimensional detection.After 500 pieces of glue-spreading image are collected and tested,it is verified that the algorithm detection speed is about 130 ms,width error of detection is controlled within 0.1mm,detection for the break of glue-spreading effectively,the position deviation is controlled within 0.2mm,which basically meets the needs of online detection.And the result can guide the robot to repair gluing.