Research On Fluorescent Magnetic Detection System Based On Deep Learning

Fluorescent magnetic particle inspection is a widely used non-destructive detection technology for surface defects of ferromagnetic materials.At present,most of the magnetic mark analysis and defect detection are performed by manual or traditional image processing.The efficiency is low,and false detection and missed detection are serious.In recent years,deep learning technology has been widely used in the field of object detection and recognition.Compared with traditional image processing methods,it generally has stronger generalization ability and higher detection accuracy.Therefore,this paper introduces deep learning technology into fluorescent magnetic particle inspection,and studies the fluorescent magnetic particle inspection system based on deep learning.In the image pre-processing part,this paper deals with the situations of motion blur,color deviation,inconspicuous image details and noise interference in the image to highlight the image details and improve the image quality.This paper first uses the Lucy Richardson algorithm to restore the motion blur of the image,and then uses the gray world method to correct the color shift in the image.For the case that the detail information of the over bright and over dark areas is not obvious in the image,this paper proposes an adaptive gamma correction algorithm processes the image,which significantly improves the local contrast of the image and enhances the image details.Finally,a bilateral filtering algorithm is used to suppress the image noise.In this paper,multiple pre-processing operations can suppress interference factors,enhance local details,and improve image quality.In the deep network model building part,a bilinear symmetric NASNet network model is constructed in this paper,which realizes the classification of the images of the defective magnetic fluorescent workpieces.In this paper,the NASNet network model is used as a feature extractor of the bilinear CNN model,which enables it to fully capture the local nuances between non-defective and defective images,and achieve accurate classification of defective workpiece images.Aiming at the problem of overfitting caused by small sample training,this paper introduces transfer learning and uses mature pre-training models to initialize network weights to reduce the risk of overfitting and shorten the training time.This paper analyzes the impact of different hyperparameters on network performance,and determines the optimal hyperparameter group to obtain the best network model.Experimental results show that the network model in this paper has higher classification accuracy on the two fluorescent magnetic particle data sets than several other commonly used classification networks,and the accuracy on both data sets is more than 99%.Finally,a hardware platform is built in this paper to collect the fluorescent magnetic particle images,and the algorithms of each part of this paper are tested.The algorithm is integrated through the software interface to realize the visualization of the processing results of each part.