Research On FPC Defect Detection Algorithm Based On Deep Learning

In recent years,with the continuous progress of human science and technology,all kinds of electronic devices appear more and more frequently in our life.And in this process,the circuit board is playing an irreplaceable role.The popularity of mobile phones,notebook computers,digital cameras and other electronic products has challenged the portability and reliability of circuit boards.The traditional printed circuit boards have been unable to meet the growing consumer needs of users.Flexible Printed Circuit(FPC)has become the best solution for miniaturization and mobility of electronic products.The material of FPC itself is fragile and the production process of FPC is pretty complex.Each process may produce different kinds of defects.If these defects are not detected in a timely manner,it will result further losses.Therefore,FPC defect detection work is particularly important.At present,the FPC industry mostly adopts the method of combining manual and instrument to complete the quality inspection.Full inspection or sampling inspection is selected according to the importance of process steps,which is inefficient and costly.Therefore,the design of automatic and intelligent FPC defect detection model is of self-evident importance to the FPC industry.With the rapid development of deep learning in recent years,many researchers have combined it with various fields and produced many breakthroughs and practical benefits.Especially in the field of computer vision,the popularity of deep neural network makes face recognition,automatic driving,industrial robots and so on become a reality.FPC has a wide variety of defects,complex circuit background and large difference in feature scale.These problems make the feature engineering of the model extremely difficult,resulting in most of the traditional defect detection algorithms based on machine vision are no longer applicable.In addition,the rapid turnover of production lines and the relative lack of defect sample data also pose challenges to the development of FPC defect detection models.Deep learning has become the key technique to solve the problem of FPC defect detection.The main work and contributions of this thesis are as follows:1.In order to solve the problems of FPC defect features,such as complex shape and variety of features,this thesis builds a new FPC defect dataset,and proposes a FPC defect detection model based on Faster R-CNN.In this thesis,by adding fusion feature pyramid,convolutional block attention module,focal loss and other sub-modules,the model can improve the ability of FPC complex defect feature information extraction.At the same time,this thesis also selects the warming-up training strategy combined with cosine-decay learning rate and ResNet101 for backbone network,so that the model converges faster and the generalization ability is stronger in the training.Compared with other detection models,the performance of our model in this thesis has significantly improved 1%in the constructed FPC defect data set,and the detection accuracy of all kinds is more than 90%and the recall rate is more than 95%.2.Aiming at the problems of lack of data and fast production line turnover in actual FPC production scenario,a defect detection model based on small sample learning is proposed in this paper.In this model,the idea of meta-learning is adopted,and the K-modal mapping module of the model is trained in stages by using the episodic training strategy,so that the model can learn task-level knowledge and avoid the occurrence of over-fitting.When there are only 1 to 10 training samples,the performance of the model presented in this paper is significantly improved compared with the traditional method,and the generalization problem of FPC defect detection model is effectively solved.