Research On PCB Bare Board Defect Detection Based On Deep Learning

The defect of printed circuit board(PCB)is one of the key factors leading to the high failure rate of electrical equipment.Therefore,the defect detection of PCB bare board is an important quality inspection technology in PCB industry,and it is of great significance to carry out accurate and real-time detection of defects.The traditional PCB defect detection method has the disadvantages of relying on template design and high cost.Because deep learning method can automatically extract image features and simplify the process of image preprocessing,it can effectively improve the accuracy and efficiency of target detection when applied to industrial production.However,the defects of PCB bare board have such characteristics as small target,many kinds and not obvious features,so it is necessary to make full use of deep neural network and feature extraction technology to identify PCB defects.In this paper,a new detection model is proposed to solve the above problems.The main work is as follows:1.In view of the shortcomings of traditional defect detection methods such as too long detection time and poor robustness,this paper adopts convolutional neural network as the dependent structure of the detection model.Based on the analysis of each layer structure of the convolutional neural network,the two-stage detection algorithm and the singlestage target detection algorithm are studied respectively.By analyzing the advantages of different algorithm structures in precision and speed,the foundation is laid for the model construction and multi-step optimization.2.In order to obtain a higher quality data set to achieve better training effect,this paper carried out the collection,cleaning and preprocessing of PCB bare board defects,formed 10,792 PCB defect data sets with high quality samples after data enhancement,and then classified and labeled the data sets.2.In order to obtain a higher quality data set to achieve better training effect,the defects of PCB bare board were firstly collected,and the invalid samples,such as fuzzy and incomplete defect areas,were cleaned through comparison and analysis.Then 10,792 PCB defect data sets with high quality samples were constructed by cropping,brightness adjustment,Angle rotation and other data enhancement methods.Finally,the data set was marked with defects by Image Labeler.3.To solve the problem of too small defect features,Faster R-CNN framework and Res Net residual block structure were adopted to solve the problem of network gradient disappearance;A multi-scale detection method based on feature pyramid was designed to detect small features,and the average accuracy of the improved model was increased by3.6 percent.In order to further improve the model performance,the feature extraction layer was improved by using deep separable convolution and inverted residual blocks to make the model structure more portable.Combining the Guided Anchoring method with the region generation network in Faster R-CNN,a large number of anchor point frames are reduced and the generated anchor point frames are closer to the target.Finally,a multistrategy detection model is constructed,which greatly improves the detection accuracy and detection speed of the model.The experimental results show that compared with the existing PCB defect detection methods,the proposed method has high accuracy,high speed and good universality,and is more in line with the requirements of industrial production.