Research On PCB Bare Board Defect Detection Algorithm Based On Deep Learning

In the era of rapid development of electronic information industry,printed circuit board(PCB),as the most basic and irreplaceable part of electronic equipment,has a large application market at home and abroad.In order to adapt to the current situation of high quality and fast PCB production in factories,and to avoid the impact on the service life and stability of electronic devices caused by unqualified PCB,it is necessary to carry out an efficient and accurate defect detection on PCB.Nowadays,the method of manual visual inspection has been gradually eliminated,while the intelligent detection method based on machine vision has been accepted by more and more PCB manufacturers because of its fast detection speed,high accuracy and simple operation steps.In this dissertation,the object detection algorithm based on deep learning is used to detect the defects in PCB,and the detection algorithm has been improved to improve the detection performance of defects.Specific research contents include:(1)Yolov3 and Yolov4 network models are used to detect defects in PCB bare board.By comparing the detection results,it is found that Yolov4 has a better detection effect than YOLOv3.In addition,the greater size of the image input into the network,the more apparent the advantage of YOLOv4 in detection accuracy,but the detection speed would be slower.(2)In order to further enhance the detection accuracy,this dissertation proposes an improved detection model based on YOLOv3.(i)The Cutmix data processing module is added into the network to enhance the generalization performance of the network model,enabling the model to pay more attention to the local characteristics of the object,and improve the performance of the model for the classification and positioning of the object.(ii)A convolutional block attention module(CBAM)is added into the network to guide the network model to pay attention to the useful feature information in PCB defect detection.(iii)A detection branch used for 4-times subsamping feature map is added to improve the YOLOv3 network structure,and to increase the number of residual blocks in the first and second residual modules of Darknet-53 network to 4 and 6 respectively,and a layer of batch normalization(BN)is added into the diffusion block of maps from different layers,enhance the network model's the performance of the small objects' defects detection performance on the PCB;(iv)The complete intersection over union(CIOU)loss is used in the detection model to accelerate the convergence speed of bounding box regression of the network during training.The experimental results show that the proposed new improved YOLOv3 network model obtains a very good detection effect in PCB defect detection,even more accurate than YOLOv4.