Research On Algorithms For Defect Detection Of Printed Circuit Boards Based On Deep Learning

As technology and the social economy evolve,electronic products are gaining an increasing array of functions,which in turn renders their circuit more intricate.This complexity makes the management and processing of circuit boards a challenging task.In the production process of PCB(Printed Circuit Board),defect detection technology is an effective method to detect defects in time.The conventional detection method misses a high rate of PCB defects because of their minute size,complexity,and variance.How to detect defects in real-time with high accuracy using a small number of training samples is a hotspot and difficulty in current research.This dissertation aims to solve this problem by utilizing techniques such as image processing and deep learning to study a PCB defect detection algorithm based on deep learning.The main research results are as follows:1.Considering the varied scales and complexities of PCB defects,we proposed a group atrous spatial pyramid pooling method to pull out features at multiple scales.A Group Atrous Spatial Pyramid Pooling-based PCB defect detection method,GASPP-SSD,was proposed utilizing the multi-scale feature extraction and SSD detection framework.This method deploys a Siamese network to derive features from both the template and test images,resulting in a difference feature map produced by subtraction of the extracted features.Following the completion of the non-linear activation,the atrous spatial pyramid pooling segmented and identified the variability in the feature map for multi-scale feature extraction.This was subsequently passed on to the SSD module for categorization of targets,box regression,and detection.Experimental trials illustrate better mean average precision with the said method as opposed to the method which uses group spatial pyramid pooling.2.A high-precision PCB defect detection method based on hybrid pooling and channel attention was proposed.To address the problem of severe information loss in traditional pooling methods,a new hybrid pooling method was proposed,and a new PCB defect detection network was constructed based on this pooling method.The network formed a dual-channel image by combining the test image and template image and used convolution and the proposed pooling method for feature extraction.The detection network was enhanced with the incorporation of a channel attention mechanism to boost detection performance.On the Deep PCB dataset,this suggested method reached an exemplary detection accuracy averaging 99.64%,showcasing the highest detection performance.3.A semi-supervised PCB defect detection method based on pretraining and singlelabel pseudo-labeling was proposed.I Considering the issue that current deep learning methodologies for PCB defect detection need many labeled training samples,a semisupervised PCB defect detection technique that involves pretraining and pseudo-labeling was suggested.The method consists of three stages.Firstly,a deep autoencoder network was added after the feature extraction backbone network to map the output back to the original input space,forming a deep self-coding network.Training was conducted unsupervised primarily using image data,with the concluding weights being employed as initial weights for the subsequent stage of training.Subsequently,a minimal number of labeled samples were utilized to instruct the detection network.Furthermore,pseudo-labels for unlabeled data were produced towards the training’s conclusion,maintaining a higher level of confidence threshold.In the end,the detection network was retrained using labeled and pseudo-labeled samples to yield the final model.Experimental outcomes on the Deep PCB dataset revealed that our suggested strategy boosted average detection accuracy by 1% to7%,in comparison to models trained with merely a few labeled samples.4.A PCB defect detection framework combining convolutional neural networks(CNNs)and Transformer models was proposed.In the backbone network of this detection framework,a residual Swin Transformer(Res Swin T)model was used to extract local detail information using Res Net and global dependency information using Swin Transformer,thereby improving multi-scale feature extraction and feature expression capabilities.In the neck of this detection framework,multi-head spatial and channel self-attention mechanisms were used to help the network focus on advantageous features in different dimensions.In the head,a multi-level cascade detector and classifier were used to improve defect detection accuracy.Extensive experiments were conducted on the PKU-MARKET-PCB and Deep PCB datasets,and through comparison with existing one-stage and two-stage detection models,the proposed DDTR model increased the F1 score by 15.42%.Numerous visualization results showed that the DDTR model had better detection performance.5.This paper proposes a lightweight PCB defect detection algorithm,BBMnet,which utilizes bounding box graph supervision.This is an anchor-free method,where a similarity feature extraction module is built to extract shared features between input template images and test image patches.By deleting these shared features,the changed areas are obtained.Finally,some convolutional blocks are used to generate bounding box graph supervision(BBM).This network structure can effectively reduce computation complexity,reduce model size,and improve detection speed.Using BBMnet,the GPP-SSD model size can be reduced to 0.6 times its original size,and the MP-SSD and HP-SSD models can be reduced to 0.4 times their original size.The real-time performance is 125 FPS,which is far superior to other methods.