Application Of Deep Learning Method In Detecting Steel Surface Defects And Chars

Steel products such as plate,strip and bar are widely used in construction,machinery,shipbuilding,automobile,military industry,household appliances and other manufacturing industries.Therefore,the production technology of steel is an important indicator to measure the level of industrial development of a country.With the development of the economy,Chinese steel companies have gradually shifted from pursuing production capacity to an intelligent,high value-added production model.Steel surface defect detection and char recognition system,as two important methods to ensure the quality of steel productions,has always drawen attentions by steel companies.At present,the existing steel surface defects detection system is mainly used for production lines with high surface quality such as the cold-rolled strip and galvanized sheet,but is rarely used for production lines with complex surface background such as medium plate and hot-rolled strip.The main reason is that the existing algorithms have limited recognition ability for complex defect samples.The application of char recognition system in steel industry starts late,the main reason is that the existing algorithm is not robust enough to recognize the char under complex surface background.In recent years,with the improvement of computational power and the establishment of large-scale datasets,deep learning method has attracted extensive attention due to its high accuracy and good generalization ability.With sufficient training samples,the recognition accuracy is far better than those of traditional algorithms.This paper mainly studies the application of deep learning method in steel surface defects detection and char recognition systems.The main research results are as follows:(1)Considering the defects with large intra-class variations,a defect classification algorithm based on multi-scale receptive field,image reconstruction and adaptive dimensionality reduction is proposed.The proposed method represents defects by fusing feature maps with different respective fields.The image reconstruction error is introduced to help the pretrained model to encode the image features when there are obvious feature differences between the defect images and the pretrained model.The feature reduction network is used to reduce the dimensionality of the features and improve the generalization ability of the network.The experimental results show that the proposed method has achieved 98.5%and 95.7%accuracy in classification of defects of plate and hot-rolled strip,respectively.Compared with directly using the pretrained model for training,it has increased by 2.2%and 2.0%respectively.(2)To solve the problems of low recall rate and inaccurate location of defects detection algorithms,this paper proposes classification priority network(CPN)and multi-group convolutional neural network(MG-CNN).Different from the traditional defect detection algorithms,the CPN classifies the steel surface image first with MG-CNN,and then selects the feature map groups according to the classification result to regress the bounding box.The MG-CNN uses mutually independent convolutional layers to separately extract each type of defect features.The experimental results show that there is an obvious spatial correspondence in the feature maps of MG-CNN,which is more conducive to defect detection and location.It has achieved good recognition results in the detection of surface defects of medium and hot-rolled plates.(3)Considering the rare occurrence of defects,a semi-supervised defect classification method called CAE-SGAN is proposed based on Convolutional Autoencoder(CAE)and Generative Adversarial Networks(GAN).The CAE is trained in a large number of unlabeled defect images to realize unsupervised feature learning.The encoder network is preserved as the discriminator of SGAN.The output layer of discriminator is modified to predict both true/false information and category of input image.The experimental results show that CAE-SGAN can make use of the unlabeled steel surface images to improve the classification accuracy of steel surface defects.In the case with insufficient training samples,CAE-SGAN method has higher classification accuracy than traditional convolutional neural network.(4)In order to realize the high-precision,low-cost char recognition system,a lightweight char recognition network based on mobilenet-v2 is proposed.The mobilenet-v2 is used as the backbone to reduce the cost of algorithm deployment.In the process of loss function calculation,threshold focal loss(TFLs)is proposed to reduce the proportion of loss of easy training samples,and to improve convergence speed of the network.Two different image augmentation algorithm,random character arrangement and image mixing,and attacking samples are introduced to improve the robustness of the char recognition network.