PCB Image Super Resolution Reconstruction And Defect Detection Base On Deep Learning

Nowadays,Printed Circuit Board(PCB)is widely used in industrial manufacturing and communication fields and plays an essential role in developing electronic information technology in the country.Therefore,the defect detection of PCB has high application and research value.During PCB production,external environmental factors may affect the collected image resolution,making it difficult for defect detection.Therefore,PCB image super-resolution reconstruction technology is crucial.Because the area of PCB defects accounts for only a tiny part of the entire board area,the human eye is difficult to identify;therefore,manual detection methods are less accurate and more costly.Although traditional image detection technology has improved the accuracy rate,the detection efficiency is still relatively low.Developing deep learning-based object detection technology has brought low-cost and efficient solutions to PCB defect detection.However,existing detection models often need help to balance accuracy and light-weighting.In summary,this paper first proposes an efficient algorithm for PCB defect super-resolution reconstruction,then improves the object detection model for light-weighting,and finally builds a PCB defect detection system using the light-weighting detection model.The main work of this paper is as follows:(1)A PCB defect image super-resolution reconstruction model based on improved SRGAN is constructed.Firstly,the generator network is constructed by using VIT and High Frequency Block modules.Secondly,VIT and Inverted Residual Block modules form the discriminator backbone network,thus effectively reducing the model space complexity while generating high-quality reconstructed images.The improved super-resolution reconstruction model performs well compared with other super-resolution reconstruction methods.All evaluation metrics are superior to SRGAN,with a peak signal-to-noise ratio of 32.27 d B,a structural similarity of 0.89,a perception index of 2.98,and a subjective average opinion score of 4.63.In addition,the discriminator parameter and model size are effectively reduced,with a reduction of 2.01 M parameters and 7.5MB model size.(2)A lightweight PCB defect detection model based on improved YOLOv5 s is constructed.Firstly,the pyramid pooling network replaces the CSP＿1 structure to construct the backbone network.Secondly,the CBAM attention mechanism is added before the 76*76prediction channel.Finally,a weighted residual branch is introduced in the neck network to promote feature fusion.In comparison experiments,the m AP.5 value of the improved model in this paper is 99.1%,the m AP.75 value is 86.5%,and the m AP.5-.95 value is 63.95%,all higher than the original YOLOv5 s.Additionally,the parameter and operation reductions of the improved model are reduced to 23% and 21% of the original models.Respectively,The speed of image inference is also improved.Compared with Faster RCNN,SSD,YOLOv5 s,and YOLOX models,the improved model has the highest degree of lightweight.In addition,the importance of the hyper-segmentation reconstruction technique for PCB defect detection is fully verified in this paper by comparing the defect detection test on the low-resolution image with the reconstructed image obtained by the super-resolution model.(3)The PCB defect detection system is designed and implemented.Regarding hardware,the Raspberry Pi 4B is used as the central controller and is equipped with an Intel Neural Compute Stick to assist in the inference of the lightweight detection model.PCB images are obtained through a high-definition camera for defect detection.The system’s software interface is designed and developed using Py Qt5 and includes real-time and offline detection modes,where IOU and confidence values can be customized for each mode.Experimental results have shown that the system can meet the application requirements for PCB defect detection.