Research On PCB Image Defect Detection Based On Deep Learning

China is a major manufacturing country in the PCB industry,the quality of PCB circuit boards is vital to the quality of enterprise products.Open circuit,short circuit,scratches,burrs and other types of defects are often several types of defects,PCB must be produced after the quality inspection,the traditional mostly manual inspection,in-line inspection,functional inspection,part of the optical or visual inspection,but the detection efficiency is low,long,high cost.With the development of image technology and artificial intelligence technology,in order to achieve high accuracy and fast detection of PCB defects,this paper proposes a deep learning-based PCB image defect detection technology,with Faster R-CNN as the core algorithm,to achieve high accuracy and fast detection of 8 types of PCB defects.The main research contents are as follows.(1)Firstly,the shortcomings of traditional PCB defect detection methods and the current research status of PCB defect detection are discussed,and the significance of PCB defect detection research based on deep learning technology is proposed;then the current research status of target detection technology based on deep learning is analysed,and the effects of different detection methods and common methods for improving target detection algorithms are discussed.(2)For common PCB defects,six types of PCB defect datasets(missing holes,mouse bites,open circuits,short circuits,burrs,residual copper)are collected,and two new types of defects,dust and scratches,are produced to form an eight types of PCB defect datasets for defect detection research.Considering that PCB defects are characterised by large scale variations and complex features that are difficult to learn,this paper builds a PCB defect detection network with the two-stage target detection algorithm Faster R-CNN as the core,and fuses multi-scale PCB defect features by building a feature pyramid network to improve the detection effect of defects of different scales.For complex samples in the dataset whose features are difficult to learn,an online difficult sample mining method is proposed to improve the detection effect on complex samples.(3)For the overlapping samples in scratch-like PCB defects,we propose to improve the detection network using non-maximal value suppression based on adaptive thresholding to enhance the detection of overlapping defect targets.To address the problem that the edge regression loss function in Faster R-CNN is insensitive and incomplete in predicting the edge regression loss,improvements using generalized crossmerge ratio loss and BalancedL1 loss function are proposed to improve the detection effect of the algorithm on PCB defects,and the detection accuracy reaches 98.71%on the test set,which can detect the defects present in PCB images with high accuracy.(4)In order to make the deep learning technology more user-friendly and practical,a PCB defect detection system based on a graphical user interface is designed to enable the training,evaluation and detection of models.To address the problem that most deep learning models are difficult to deploy on devices with low computing power due to their large size,a lightweight Faster R-CNN algorithm based on MobileNet v2 is proposed,which improves the detection speed and can detect an average of 16 images per second,reducing model size and facilitating the deployment of PCB defect detection networks on mobile devices.Figure[45]Table[12]Reference[74]...