A Research And Application Of Magnetic Ring Appearance Defect Detection Method Based On Deep Learning

With the rapid increase in demand for electronic products,magnetic materials are the most indispensable components of electronic components,such as capacitors,inductors,and ICs.Therefore,the market demand for the appearance defect detection of magnetic materials has increased with the expansion of the electronic product market.increase.In the field of magnetic material appearance defect detection,the traditional detection method basically relies on human naked eye recognition,with lower efficiency,higher cost,and high false detection rate.In recent years,machine vision has also been used to conduct visual inspection,and it has gradually begun to penetrate the entire field.However,machine vision also has shortcomings,a single scene,expensive equipment,and difficult algorithm design.In this paper,in view of the above existing problems in the field of magnetic material appearance defect detection,combined with recent developments in deep learning,the research and application of deep learning-based magnetic material appearance defect detection are proposed.The focus is on the convolutional neural network,especially the YOLO algorithm in target detection,and the experimental analysis is carried out on the existing sample data.The experimental results prove the feasibility of the method proposed in this paper.Aiming at the problems of the adopted YOLO algorithm on the sample data of this experiment,two improvement measures were proposed.The first direction is to add the CBAM embeddable network to the experiment,increase the attention model on the spatial features and channel features,and achieve a mAP increase of about 1% without significantly increasing the network parameters and model volume.The second direction is to use the scaling factor in the batch normalization(BN)layer,add a sparse factor to sparsely train the model,and then perform pruning operations on the model according to the pruning ratio.The experiment proves that the model does not change when the mAP is basically unchanged.The parameters and model volume can be reduced to 1/6 of the original,and the forward inference time is reduced to 1/3 of the original.This helps to deploy network models in embedded devices with limited GPU computing resources and storage resources.Service architecture for edge computing.