Research On Detection Algorithm For Surface Defects Of Hot Rolled Strip Steel Based On Convolution Neural Network

The surface quality of hot rolled strip is easily affected by production environment,mechanical equipment and processing technology,and therefore it often creates defects such as rolled-in scale and scratch on its surface,which affects the product performance.To solve this problem,from the perspective of product production intelligence,detection algorithm of the hot strip surface quality monitoring system by using the deep learning theoretical knowledge in the current machine learning field is studied in this work,and the classification and location of hot strip surface defects are also explored.The specific contents and results include the following aspects:1)The morphological characteristics of surface defects based on hot rolled strip are analyzed,and the optimization function and structure in convolution neural network model VGG16 are improved.Therefore,a hot rolled strip surface defect classification method based on optimized convolution neural network model is constructed,which improves the accuracy and speed of hot rolled strip surface defect classification.2)Designed a few-shot classification model based on deep learning,and the feature fusion strategy is involved to make up for the lack of defect sample information and therefore improve the accuracy of surface defect of hot rolled strip.In the actual production process,the defect incidence rate of hot-rolled strip steel is lower,so it is time-consuming and labor-consuming to collect a large number of defect samples and label their defect categories.In the simulation of human learning paradigm,the prior knowledge learned by CNN in other auxiliary data sets is transferred to the identification task of hot-rolled strip defects,and a small number of labeled hot-rolled defect samples are used to fine-tune the classifier parameters of the trained model.At the same time,the feature fusion strategy is designed to make up for the lack of defect sample information under the condition of only a small number(only one)of hot-rolled strip with label.The results show that the small sample learning method is effective for the defect classification of hot-rolled strip steel without training samples,and the feature fusion method can supplement the sample information.3)Established a real-time defect detection model based on skip layer connection and feature pyramid fusion module with high-precision to speed up the detection speed.The surface defect classification algorithm of hot rolled strip can only give one defect category to a defect image.In this paper,an end-to-end defect target detection algorithm named RDN is designed to detect the defects on the surface of hot rolled strip steel when the defect image may contain two or more types of defects.Defect target detection technology can identify multiple defects in an image and locate their positions.Considering the application of detection algorithm in actual production,this paper uses a lightweight and structured network backbone for defect feature extraction,and designs a jump-layer connection module and a pyramid feature fusion module to adapt to different defect image characteristics.The experimental results show that the RDN method has a good performance in the surface defect detection of hot rolled strip steel,and the relationship between detection speed and accuracy is well balanced.4)Established a compression classification model based on channel pruning.When the surface defect detection algorithm of hot rolled strip is applied to the actual industrial production,not only its detection performance should be considered,but also its required hardware cost should be considered.The model compression method can compress the test model,which makes it possible for the test model to be better applied to low-end equipment.In this paper,a hardware-friendly structured pruning method is used to prune the model channel to reduce the number of model parameters.At the same time,in order not to introduce extra computation,the scale transformation factor contained in the batch normalization layer in the network model is used to evaluate the channel importance.After fine-tuning the compression model,the results show that the model compression method we used can greatly reduce the calculation amount of the hot rolled strip detection algorithm without significantly reducing the model performance.5)The defect identification performance of the above method in the actual hot strip production is verified.The CNN based classification algorithm and Few-shot learning methods are respectively applied to the construction of real-time classifier in hot-rolled strip monitoring system.Through the classification and identification of various defects collected on the production site and performance evaluation,the results prove that the classification algorithm designed in this paper is effective in the identification of actual production.