Real-time PCBs Defect Detection Based On Deep Learning System Research

Printed Circuit Board(PCB)is the core component of electronic products,and its quality directly affects the performance and reliability of the entire product.However,PCB defects are inevitable in the production process,which leads to the failure of electronic products to work normally and even safety hazards.Therefore,the automatic detection of circuit board defects is particularly important in improving the yield of enterprises,ensuring efficient productivity,and enhancing brand reputation.It is of practical significance to promote China ’s competitiveness in intelligent manufacturing and industrial digitization.At present,there are many problems in PCB defect detection methods,such as high subjectivity of manual detection,inability to adapt to the high initial cost of laser imaging in complex circuit detection and optical detection systems,many problems in use and maintenance,and complex system circuits.In order to reduce the cost and maintenance problems of PCB defect detection system,the real-time,automation and convenience requirements of embedded detection system are realized.Therefore,this paper proposes Centernet-SPMFF and YOLO＿AD networks based on deep learning algorithms.Through experimental comparison,the optimal improved model uses MVC architecture for hardware migration,and finally builds a real-time online PCB defect detection system prototype.The specific work is as follows :(1)The research significance and current situation of PCB defect detection system are introduced,and the advantages and disadvantages of current optical system for PCB defect detection technology are analyzed.Combined with the application of deep learning target detection,different types of Centernet and YOLOv5 network models are selected as the basic network models for PCB defect testing based on the characteristics of prior boxes and network parameters.(2)The problem of hardware selection and construction in PCB defect detection system is solved by TX2 hardware platform.A complete embedded neural network system for defect detection is implemented by MVC(Model View Controller)architecture.The hardware network migration deployment is studied and the cost,GPU performance,running memory and calculation parameters are analyzed.The system is developed and deployed with TX2 using MVC architecture.(3)Aiming at the problems of high cost and single detection of PCB defect detection system,an improved Centernet and YOLOv5 network model is proposed,and the hole residual convolution structure,Series and Parallel Multi-scale Feature Fusion(SPMFF)and Ghost module are used to optimize the network structure,improve the detection efficiency and compress the model,so as to realize the contrast migration of the model.The centernet network only changes the receptive field of the defect feature without increasing the amount of computation,and then combines SPMFF to obtain high-resolution and smalltarget defect feature information;YOLOv5 uses the lightweight Ghost module as the backbone feature network,integrates the channel attention mechanism,and introduces the hole residual network,thus avoiding the convolution calculation of the network model and improving the network calculation efficiency.(4)By comparing the improved Centernet-SPMFF network and YOLO＿AD network through ablation experiments,the YOLO＿AD model was finally selected for PCB defect detection system from the comparison of model capacity,detection speed and detection accuracy.The former improves the detection accuracy and reduces the model capacity.The latter network capacity and network parameters are reduced by about 64 %compared with the original network,which greatly improves the target detection speed of the network.(5)Finally,a real-time detection prototype integrating image acquisition,model recognition and detection result visualization is established.The intelligent detection system device detects various defects of PCB with a recognition rate of 92.80 % and a detection speed of 30 FPS.