Research On Defect Detection Of Industrial Steel Based On Structural Re-Parameterization

In recent years,because of the impact of the epidemic,the demand for steel in the mills has decreased.The arrival of the post-epidemic era has led to a gradual increase in the demand for steel in factories.However,China is a large steel-producing country and its productivity ranks first in the world,as productivity increases,more and more defective steel products are also being produced,these defects will bring many safety hazards.Many steel manufacturing companies in China have experienced a high defect rate in the steel produced,resulting in millions of assets being lost each year so how to control the rate of steel defects is also very important.With the development of deep learning technology in recent years,many researchers are applying deep learning in steel defect detection research,which improves the effect and reliability of steel defect detection techniques.Currently,however,the corresponding sensing technology is not very mature and can not reach the trade-off between accuracy and speed.For the purposes of this thesis,targeting steel defects in the industrial steel production process,based on the requirements of industrial deployment,by comprehensively considering the accuracy and speed,and by improving on the basis of YOLOv5,combined with the structural re-parameterization method,the defect detection model of industrial Re-YOLOv5 steel is built.The main content is as follows:(1)Structural re-parametrization based on training and inferring and understanding the coupling architecture is investigated.We build a series of structures for training,and transform their parameters to another set of parameters for reasoning,in order for a series of structures for training to be transformed into another series of structures for reasoning.We use multi-branch structure in training and add the residual modulus to pay attention to precision and to save memory;in reasoning,we use the single branch structure to pay attention to the speedup,so that the model can improve speedup without losing accuracy.To test the effectiveness of the structural reparameterization,we perform image classification experiments on the ImageNet,NEU-DET,and GC10-DET data sets.(2)The Re-YOLOv5 is a proposed design.In addition to deepening network performance,the use of the one-stage ’s YOLOv5 algorithm,embedded in the RepVGG module,can also improve reasoning speed,as well as achieve the trade-off between accuracy and speed.The Neck layer and Head layer of YOLOv5 are merged into Head layer,which is used as prediction,and RepVGG module and convolution are added to output the prediction result.The RepVGG module adopts structural re-parameterisation method and multi-branch architecture during the training phase,which may allow the model to better learn features and improve accuracy.During the reasoning,an equivalent transformation is performed,and the multi-branch structure is re-parameterized to a single-branch structure in order to realize the speedup of the reasoning.(3)The introduction of the enhanced SPP* layer in the network model can increase the receptive field,enhancing the ability of network feature fusion,as well as improving detection accuracy for different target types;it is useful to upgrade the CSP component to the CCBL component in order to deepen the model and better obtain receptive field information.The Backbone composed by it is used for feature extraction,which can improve the speed of reasoning of the model and improve the accuracy of the detection at the same time.(4)We propose a new approach to feature extraction.Tested on the open NEU-DET steel defect data set,the detection accuracy of the model Re-YOLOv5 is 77.8%,which is slightly smaller than that of YOLOv5 l,but the number of parameters is reduced by a factor of 7.65,and the model also reduces by 80.6MB.Compared to the YOLOv5 s,Re-YOLOv5 ’s single frame image processing,the time consuming is only increased by 3.4ms,but its accuracy is improved by 6.0%,and the corresponding number of parameters and models is reduced.We have also tested on more complex GC10-DET data sets,and the results show that the accuracy of Re-YOLOv5 is still1.4% better than that of YOLOv5 s,which is a reflection of the benefits of the model proposed in this study.The model also takes up less memory and is easily deployed to industrial devices.