Research And Application Of Automatic Recognition Technology For Circuit Board Components Based On Deep Learning

This paper conducts in-depth research on the automatic recognition technology of circuit board components based on deep learning,and it is used to detect missing soldering,wrong soldering,wrong insertion,and missing insertion in PCB production.This paper designs and implements a PCB inspection system,and it consists of four modules: image stitching,component detection,text recognition,and PCB defect analysis,which can detect defects such as missing soldering,wrong soldering,wrong insertion,and missing insertion in PCB,effectively improving PCB production qualification.In order to achieve the goal of rapid defect detection,this paper conducts in-depth research on PCB defect detection based on deep learning,involving object detection and text recognition technology.Objection detection is used for component detection,and text recognition is used to recognize texts on components.This paper proposes a PCB component detection based on lightweight network algorithm for the problem of component detection on circuit boards.For the problem of printed text recognition on components,this paper proposes a lightweight text recognition algorithm for printed on components.It effectively improve the speed of component detection,increase the accuracy of text recognition and greatly reduce the amount of parameters of the entire model.The main research contents of this article are as follows:(1)This article first briefly describes the current status of PCB defect detection and text recognition in recent years,and analyzes representative algorithms in object detection,traditional algorithms and algorithms based on deep learning in text recognition.(2)This paper proposes a PCB component detection algorithm based on lightweight networks.This algorithm first uses the modified PeleeNet network for feature extraction,then uses the Region proposal network(RPN)to obtain the object proposal boxes,and uses Context-Aware ROI Pooling to normalize the object proposal boxes.Finally,RCNN is used to locate and classify the object.(3)This paper also proposes a lightweight text recognition algorithm for printed on components.This algorithm first corrects the deformed text through the Spatial transformer network(STN),then uses the modified Pelee Net network to extract feature sequences,then combines the Dense Block module at the recurrent neural network to extract the label sequences,and finally the prediction of the label is achieved through the attention mechanism model.(4)In the end,this paper designs and implements a PCB detection system.This system combines the two algorithms proposed above to achieve missing soldering,wrong soldering,wrong insertion,and missing insertion on PCB.This system mainly includes four functional modules: image stitching,component detection,text recognition,and PCB defect analysis.