Research On PCB Surface Defect Detection Model Based On Faster R-CNN

In recent years,with the continuous improvement of computer computing power,deep learning has gradually become a key technology of the intelligent industry in the new era,and it has achieved remarkable results in the fields of face recognition,natural language processing,weather forecasting,and financial analysis.The printed circuit board(PCB)is a basic electronic component,and its quality is related to the performance of the entire electronic device.Therefore,it is extremely important to detect defects on the PCB.This article will detect PCB surface defects in computer vision.The types of PCB surface defects are complex.The traditional defect detection efficiency is low and is greatly affected by human factors.The production cost is high and the detection accuracy cannot be guaranteed.Deep learning can extract feature information from the bottom to the top of the data,so it is widely used in product defect detection.Compared with other commonly used detection models,Faster R-CNN has higher accuracy and faster speed,but its backbone network VGG has a complex structure,and the calculation speed still cannot meet the actual production requirements;the regional candidate network(RPN)in Faster R-CNN cannot Making full use of semantic information leads to the problem of deep network degradation.Therefore,this paper uses Res Net instead of VGG as the backbone network,and introduces the feature pyramid network(FPN)to solve the problems of gradient dispersion and deep network degradation,so as to perform the Faster R-CNN model.Improvements have greatly improved the detection speed and accuracy.The main research contents of this paper are summarized as follows:(1)Aiming at the shortcomings of VGG parameters and slow calculation speed,this paper uses Res Net instead of VGG as the Faster R-CNN backbone network.The Res Net network adds a short-circuit connection before the input and the activation function of the second layer.The complexity is lower than that of VGG.The number of layers is deeper but there is no degree gradient dispersion phenomenon,which can effectively solve the problem of deep network degradation.Aiming at the problem that semantic information is basically disappeared from convolution pooling to the last layer during small target detection,FPN can merge high-resolution shallow feature information and deep feature information with rich semantic information through upper and lower structures and horizontal connections.To generate strong semantic feature information at all scales,thereby effectively solving the problem of multi-scale detection,and at the same time reducing the computational cost.(2)Compare the commonly used deep learning models.Due to the small amount of data,the data is preprocessed first,and the brightness is changed by flipping and clipping to expand the data to meet the data set required for deep learning.Experiments confirm that Faster RThe CNN model works best in PCB defect detection.Then compare the model before and after improvement,enhance the data of the expanded data set and conduct experiments.By comparing the improved model in this paper with the original model,the experimental results show that the improved model in this paper is significantly better than the original model in the detection of PCB surface defects,with an accuracy rate of 91.8% and an average detection speed of 1.45 seconds per image,which can effectively improve Detection accuracy and speed.