Research On Surface Defect Detection Method Of YOLOv5 Metal Sheet By Integrating Attention Mechanism

Metal plate is one of the important products in the metallurgical and manufacturing industries.As an indispensable raw material in the industrial fields of automobile manufacturing,mechanical processing,chemical materials,aviation and shipbuilding,it is widely used in the production and manufacturing of modern life.During the production process of metal plates,due to external process factors such as billets and rolling processes,surface defects such as black spots,ink,and welds often occur.The surface defects of metal plates not only affect the appearance of products,but more importantly,they can affect various physical properties of metal plates to varying degrees,such as their wear resistance and fatigue resistance.In serious cases,they can cause incalculable economic losses and adverse social impact on production enterprises.However,due to the complexity,poor robustness,and easy to miss detection,the current general target detection methods have the problems of low classification rate,low recognition rate of single defect targets,large memory consumption,and low real-time performance,which are generally difficult to meet the requirements of industrial applications.In order to address the above challenges,this article conducts the following research:1.A new metal sheet surface defect detection network(MSSDN)is proposed.The network mainly integrates attention mechanisms,and its core part is composed of two parts.One is a backbone network based on local two-way self attention,which is responsible for extracting basic image features,determining which information the network integrates,and which redundant information is discarded.The other part is a shapeable detector head that allows the receptive field of the convolutional kernel to undergo adaptive changes based on image content,paying attention to more information.2.Focusing on model lightweight is crucial for industrial networks with high real-time and memory requirements to reduce their storage and computing costs.The MSSDNN network in the previous chapter has a high accuracy and recall rate,but due to the limitations of model size and floating point computation,it does not work well on devices with low computational power.In this paper,model quantization is used to reduce unnecessary floating-point calculations in MSSDDN.Dynamic Bit Width Selection(DBWS)is used to quantify the trained model.Based on this,an improvement is made to the activation function of the algorithm based on dynamic interception of activation values,which simplifies the calculation times of the network and the volume of the model.The improved detection method was validated on the metal plate surface defect detection data sets NEU-DET and GC10-DET.The results show that the average accuracy of the full category is improved by 7% and 4.2% respectively,and the volume compression ratios of the quantized model are 3.667(from 331 Mb to 90 Mb)and 3.609,respectively.The acceleration ratios under single thread are 1.32 and 1.26,respectively.Moreover,the processing time for each four images of the model under four thread concurrent conditions is only 7.25 milliseconds.In addition,this method is also superior to existing general detectors in terms of detection speed and recall,and can better meet the needs of industrial scenarios.