Research On PCB Defect Recognition Based On Image Autonomous Learning

With the explosive development of electronic products in recent years,printed circuit boards tend to be miniaturized and high-density design forms,but circuit board defects occur frequently,which makes people put forward higher requirements for the quality of circuit boards.The traditional circuit board defect detection is mainly realized by manual visual inspection,which will cost a lot of cost and the accuracy of detection is also very low.Therefore,the defect identification technology based on printed circuit board needs more efficient and faster methods.Firstly,the paper studies the defect identification of circuit board by using image processing technology,and accurately registers the standard image and the image to be tested,and makes the difference operation to get the position of the circuit board defect.Then,a method based on edge detection is proposed to classify different characteristics of different defects.Then,the experiment adopts the method of circuit board defect identification based on convolutional neural network.By comparing several more efficient target detection networks,we confirm that the open source single stage target detection network yolov3 network is used.The network can locate and classify the defects of circuit board at the same time.The experimental results show that YOLOv3 network can be used in the accuracy or speed of edge detection algorithm The quality of the system has been improved,but there is still a problem of insufficient accuracy of defect identification.Therefore,in order to make the network model be able to recognize circuit board defects autonomously,this study proposes a D-YOLOv3 network based on improved YOLOv3.First of all,in view of the idea of Dense Net,this paper introduces Dense Block module in Dark Net-53,the backbone of YOLOv3,to enhance the reuse and propagation of network features,and solve the problem of gradient disappearance of deep network to a certain extent.At the same time,a prediction layer is added to the improved network to make full use of the shallow information.In addition,this paper improves the K-means clustering algorithm.When calculating the distance of centroids of different sizes,the corresponding weights are added to constrain them to smaller anchor points,so as to improve the small object detection suitable for circuit board defects.In the loss function optimization,this paper proposes to replace the original IOU loss with GIOU loss.The experimental results show that the D-YOLOv3 network proposed in this paper has a great improvement compared with the edge detection algorithm and the YOLOv3 network.Secondly,because the introduction of the core module of Dense Net can reduce the parameters of the model to a certain extent,the speed of the D-YOLOv3 network model also shows a small improvement.