PCB Defect Detection Based On DcMsNet

PCB defect detection is a key stage in the production line of the electronics manufacturing industry to ensure product quality.With the improvement of product quality,mobile phones,home appliances,smart cars,space shuttles and so on have put forward higher and higher requirements for the reliability and robustness of PCB products.However,the PCB defect detection currently used in the industry still has problems of unsound features,unbalanced samples,small target detection,and noisy samples.PCB defect detection usually uses classification methods based on manual features,which can only be designed based on human experience,often with simple patterns,furthermore,the extracted features do not have strong characterization and robustness.Secondly,due to the extremely imbalanced categories of PCB defect detection tasks,the deep network focuses on the categories with a large number of samples in the learning process,and cannot fully learn the characteristics of small sample categories,resulting in poor recognition performance of these samples.At the same time,there are small,irregular curves and shape detection in the task,which will also reduce the performance of defect detection algorithms.Finally,there are a large number of noise samples in PCB data sets.The above four problems all reduce the comprehensive performance of the PCB defect detection algorithm in the actual industrial environment.Therefore,in response to the above problems,this paper puts forward a variety of PCB defect detection algorithms based on deep neural networks for actual industrial scenarios,and conducts sufficient experiments to verify them.Compared with existing systems,the main contributions of this paper are:Firstly,in response to the extreme imbalance of data set categories in the production environment,this paper proposes a cyclic network composed of two sets of one-way adversarial generative networks,which are used to generate defective categories with a lack of sample size,alleviate the imbalance between categories,and enrich data set features.Secondly,in response to the problem of various sizes and shapes of objects to be identified in the production environment,this paper proposes a new deep network Dc Ms Net embedded with deformable convolution and multi-step classification modules,in which the residual unit performs structural transformation,after adjusting the order of different components in residual modules,to improve the ability of the residual network to extract features,DcMsNet's backbone is ResNet101,the deformable convolution in the DcMsNet is used to adapt to different shapes of targets.The classification task of different size targets is realized based on multi-stage learning of auxiliary classifiers.Thirdly,in response to the problem of noisy data set in the production environment,this paper proposes a customized loss function with label smoothing to complete the noisy classification learning task,and reduce interference of noise samples to the network during the training process.In order to verify the effectiveness of the method proposed in this paper,this paper built a data set of PCB defect detection with different shapes and sizes,then achieved an accuracy of 97.3% on this data set.The experimental result shows that the method proposed in this paper has high precision and recall rates for all six defect categories.Compared with other existing methods,this method has better classification performance.