Research On Character Recognition Of PCB Components Based On 3D Vision

With the rapid development of electronic information technology,PCB is more and more widely used.In order to detect it accurately and efficiently,AOI has been gradually developed and has been applied to practical production.Nowadays,the AOI technology mainly adopts 2D visual detection methods,but because of the different printing methods of the components surface characters,the components characters are concave and convex,the colors and sizes are diverse,and the 2D character recognition system has low robustness and insufficient accuracy.The height information in the 3D point cloud of PCB(3D-PCB)data can effectively enhance the character information in the 2D image(2D-PCB)and improve the contrast between the character and the background.Therefore,it is of great significance to develop a character recognition system based on 3D vision to improve the accuracy of character recognition in AOI system.This thesis designs a character recognition system for PCB components based on 3D vision.The system proposes 2D/3D registration and splicing methods based on data calibration and marking 3D-PCB from the point of view of fusion of 2D images and 3D-PCB data,and completes the fusion of 2D images and 3D-PCB data and the splicing of 3D-PCB images.The surface 3D the collected components is uneven due to the fluctuation of 3D line sweep.In view of this problem,a character enhancement algorithm based on RANSAC plane fitting is designed to enhance the characters in 2D images by using 3D information.Besides,the character segmentation algorithm in complex environment is proposed to complete the character segmentation of components better,and the character recognition is completed by using MobileNetV3.The main elements are as follows:(1)General scheme design of PCB component character recognition system.Firstly,the technical requirements of character recognition of PCB components are analyzed,and the hardware selection and layout designs of character recognition system are completed.Secondly,in order to collect 2D images and 3D data,the operation flow of the acquisition system is designed.Finally,according to the character characteristics of components,the overall algorithm flow of the system is preliminarily designed.(2)2D/3D image registration and splicing based on data calibration and marking points.Using the rich texture information in the PCB image to be stitched,a FAST feature and RANSAC registration method is proposed to splice multiple local images.Secondly,aiming at the problem that 2D-PCB images and 3D-PCB data from different sensors can not be directly corresponding,an information normalization algorithm based on data calibration is proposed to realize the coarse registration 2D-PCB images and 3D-PCB data.finally,a mark-point-based 2D/3D registration and splicing algorithm is proposed to obtain D-PCB conversion relationship between 2D-PCB images and 3D-PCB data.The matching alignment of the 2D-PCB is completed.The accurate registration of 2D-PCB images and 3D-PCB data is further realized,and the integrity of components and their characters in 3 images is guaranteed.Experimental results show that the proposed registration and splicing method can obtain complete 3D-PCB images.(3)Character localization and segmentation based on 3D data.First of all,in order to select the components to be detected from a large number of components,the through filter is used to remove the area in the PCB board without detection,and the mask of the components to be detected is generated.Secondly,in view of the noise fluctuation,3D data on the surface of PCB components and the inability to obtain accurate projection datum,a character image enhancement algorithm RANSAC plane fitting is proposed.Ensure that the 3D data of the components can effectively enhance the characters in the 2D-PCB image.The character region is accurately located by combining MSER and morphological method according to the interference of PCB components by external dirt.Finally,the character segmentation algorithm in complex environment is designed to complete the character segmentation of all kinds of PCB components.The experimental results show that the character enhancement algorithm proposed in this thesis can effectively enhance 2D character images,and the character segmentation algorithm designed in this thesis can accurately segment PCB component characters.(4)Character recognition of PCB components based on MobileNetV3.Aiming at the interference information of the segmented image,the method of connected domain analysis is used to remove the interference information,and a clean PCB component character segmentation image is selected to improve the character recognition rate.moreover,the exclusive data set of PCB component characters is constructed and the MobileNetV3 neural network is trained using the training set.PCB component character recognition is completed using the updated network parameters.The comparison test with SVM shows that the MobileNetV3 based PCB component character recognition method network can recognize the component character more accurately and quickly with the same test set.To sum up,according to the technical requirements of character recognition of PCB components,the hardware system of character acquisition of PCB components is built,including 2D acquisition modules and 3D acquisition modules.The character recognition algorithm based on 3D vision is designed,and the PCB component character recognition software system based on 3D vision is developed based on the software platform and modularization idea.It mainly includes user login module,character recognition module,data management module and training set construction module.Field experiments show that the average accuracy of the character recognition system designed in this thesis is 98.28% which basically meets the technical requirements of character recognition of PCB components.