Research And Application Of Surface Defect Detection Of Metal Sheet And Strip Based On Deep Convolutional Neural Network

Sheet metal strip is one of the main products of the metallurgical industry and is widely used in pillar industries such as home appliances,military and chemical industries.Surface quality is one of the key quality factors of sheet metal strips.Traditional machine vision-based surface defect detection algorithms have disadvantages such as insufficient detection results,slow detection speed,poor promotion performance and poor adaptability.Therefore,it is necessary to study intelligent automatic defect detection algorithms to improve the efficiency and accuracy of defect detection.This paper takes the surface defects of metal plates and strips as the research object,and applies Deep Convolutional Neural Network(DCNN)technology in the field of deep learning to the problem of image detection of surface defects of metal plates and strips.The algorithm research is mainly carried out from the aspects of surface defect image classification and target detection,and the research results are verified through system design and development.The main research contents of the paper are as follows:(1)Aiming at the difficulty of surface defect feature extraction and the limitation of data sample collection,a defect image classification algorithm d DCGAN-DCNN combined with improved Deep Convolutional Generative Adversarial Networks(DCGAN)and DCNN is proposed.The algorithm first improves the network structure of the DCGAN generator and discriminator,and trains it with NEU-CLS training samples and random noise on the hot rolled steel strip surface defect classification dataset to obtain an improved DCGAN high-resolution defect sample generation model.On this basis,the generated samples and training samples are combined as a new training set,and a VGG-16 transfer learning model is constructed according to the research object to perform the defect classification task,and the validity of the generated samples is tested.The comparative experimental results show that the defect features automatically extracted by the algorithm have strong discrimination ability,and the generated samples can improve the accuracy of defect classification,the classification accuracy reaches 99.07 %,and has good noise immunity and generalization.(2)Aiming at the problem of the diversity,complexity,and randomness of the surface defects of the sheet and strip,which makes it difficult to quickly locate and accurately identify them,a multi-level feature Faster R-CNN-based defect target detection algorithm Network,DDN).The algorithm uses a Multilevel-Feature Fusion Network(MFN)to fuse the various feature maps extracted by VGG-16 in Faster R-CNN to obtain a fused feature map with rich location and semantic information.The subsequent network is based on this The feature map is fused for classification and border regression to finally determine the category and location of the defect.Using the NEU-DET and KMUST-DET datasets of steel strip and copper plate surface defect detection to evaluate the performance of the proposed algorithm,the comparison experimental results show that the proposed DNN can quickly and accurately detect various types of defects with different scales.It has better detection accuracy without losing too much detection time.The average detection time is 129.65 / 153.17 ms,and the Mean Average Precision(m AP)is 86.13 % / 92.54 %.(3)According to the functional requirements of the detection system,apply(1)(2)research algorithms to the detection system,and design and develop an offline intelligent detection system for strip surface defects based on DCNN.The system is based on the B / S architecture,with j Query Easy UI as the UI tool in the foreground,Django as the web application framework in the background,Postgre SQL as the DBMS,and Tensor Flow and Open CV as the visual algorithm development kit.Adopt Ajax technology to realize front-back background data interaction,and use GPU and cu DNN acceleration framework for network training.It implements functions such as user rights management,defect sample database construction,defect image preview,model training,model viewing,defect classification,defect target detection,and inspection result query.The operation results show that the designed system has good interactivity and user-friendliness,can complete the prescribed tasks,and meets the needs of users.