Research On Detection Method Of Magnetic Tile Surface Defects Based On Fully Convolutional Network

Magnetic tile surface defect and recognition is an essential part of industrial production.In the production of magnetic materials,accurate selection of defective magnetic tile products can increase the performance of the product.In large-scale production of magnetic materials,a lot of manpower is required to screen out the defective magnetic tiles during the production process.To saving labor cost and ensuring the quality of the magnetic tile,it is very important to realize automatic detection and identification of magnetic tile defects.Based on the research of domestic and foreign researchers on the surface magnetic tile defect detection and target detection,this paper proposes two surface defect detection methods for magnetic tiles.The main work is as follows:(1)Aiming at the problem of the small amount of collected magnetic tile data,the conditional generative adversarial network and data enhancement methods are used to amplify the magnetic tile data,so that the sample size of each type of data is in a balanced value,and the occurrence of overfitting is reduced;Pixel segmentation method is used to separate the magnetic tile defect characteristics from the background,so as to improve the contrast of the magnetic tile defect itself and accelerate the speed of defect identification by the model.(2)A magnetic tile surface defect detection method based on Multi-scale YOLOv3 network is proposed.Because it is unreasonable to apply the anchor box in the YOLOv3 network to the magnetic tile data set,the K-means II algorithm is used to re-cluster the magnetic tile data set to obtain a new anchor box,and the defect target location in the magnetic tile data set can be accurately located;when training the deep neural network model,this paper merges the BN layer with the convolutional layer to reduce the calculation amount of the model to improve the speed of forward inference;to improve the detection rate of the small objects,a multi-scale mutual fusion network structure is proposed.By constructing a four-scale network architecture,low-resolution features(high-level features)are integrated into high-resolution features,which can improve the resolution of features and enhance the boundary information of the objects.The results show that the average recognition accuracy of Multi-scale YOLOv3 networkfor magnetic tile defect detection and recognition is95%.In terms of the speed of detection and recognition,the FPS of the Multi-scale YOLOv3 network is 29.88.(3)A magnetic tile surface defect detection method based on SE-Res Net Faster RCNN network is proposed.Because the main network used by Faster-RCNN is VGG16,the network layer is only 16 layers,and the deep semantic information of the network cannot be obtained when extracting small target features.Therefore,this paper uses Res Net-50 with deeper layers as the main network of Faster RCNN;The attention mechanism is combined with Res Net-50 to suppress unimportant background information while enhancing the important information in feature extraction of magnetic tile defects,so that the network can detect magnetic tile defects more accurately;Soft-NMS is used to screen the prediction candidate boxes generated by the network to increase the detection accuracy of the algorithm.The results show that the average recognition accuracy of SE-Res Net Faster RCNN network for magnetic tile defect detection and recognition is 89.2 %.It can be seen from the experimental results that the two methods proposed in this paper can better identify the surface defects of the magnetic tile,indicating that the two models proposed in this paper have certain robustness and meet the requirements of real-time in industrial production.