Research On Defect Detection Technology Of PCB Bare Board Based On Deep Learning

With the rapid development of China's PCB industry,PCB bare board defect detection has a high research and application value as an important link in PCB production.Aiming at the problems in the existing PCB bare board defect detection technology,this thesis proposes a deep learning based PCB bare board defect detection technology.First,this thesis discusses the research background of PCB defect detection and the current research status at home and abroad,and briefly introduces the related algorithms and detection processes in the PCB defect detection technology that is more commonly used based on AOI system and machine vision,and analyzes its limitations and deficiencies.And starting with neural networks and deep learning,this thesis introduces the relevant algorithm model based on target detection,analyzes the feasibility of applying the target detection network to PCB defect detection,and its superiority compared with traditional PCB defect detection algorithms.Secondly,this thesis refers to the structure framework of the Faster R-CNN network model,and optimizes and adjusts the difference in data between the bare PCB defect detection task and the traditional target detection task.This thesis compares different convolutional neural networks,selects ResNet-101 as the feature extraction network,and fuses multi-scale features through bottom-up pathway,top-down pathway,and lateral connections to build a feature pyramid to improve the detection effect of the network on defects of different sizes.This thesis combine RoI Align optimization algorithm in the RoI pooling layer to reduce the loss of feature accuracy,and add batch normalization layers to the network to increase the training speed of the network and effectively suppress the occurrence of gradient dispersion and over-fitting.After that,in this thesis,the convolutional neural network cannot distinguish the noise features from the target features when extracting features.In the case of serious noise interference and blurred boundaries between the target and the background,it is prone to miss detection and false detection.The noise suppression network of the expanded convolution and residual module.Compared with traditional HLM and BM3 D denoising algorithms,the effectiveness of noise suppression network denoising is proved.Finally,this thesis selects six common PCB bare board defects as the experimental defects,including: missing hole,mouse bite,open circuit,short,spur,and spurious copper.A PCB bare board defect data set is used to train and test the network model.During the training process,the anchor size is fine-tuned on the basis of K-means clustering through statistical data distribution of the training data,and the learning rate and weight attenuation coefficient during network training are selected through the 1cycle strategy to reduce the difficulty of thetask and save the network Training time,reduce the degree of over-fitting and improve detection accuracy.The experiments show that the deep learning-based PCB bare board defect detection method proposed in this thesis is feasible and effective,and has certain reference value for applying deep learning algorithms to PCB defect detection.