Intelligent Detection Technology Of Plate Surface Defect Based On Machine Vision

With the continuous development of the national manufacturing industry,sheet metal has been widely used in various fields of industrialization,with a huge demand.However,due to the production process,the environment and other reasons,various defects will affect the use of the plate.Artificial visual inspection is highly subjective,time consuming,and laborious.Traditional image processing algorithms require selfmade templates with low accuracy and slow speed.Therefore,new methods are needed to solve the problem of detecting surface defects in sheet metal.In this thesis,the aluminum dataset was taken as the research object,and the YOLOv5 algorithm in deep learning was used to detect the surface defects of aluminum materials.To do comparative experiments and analyze the experimental results,the main work completed in this thesis was as follows:(1)This thesis compares the advantages and disadvantages of different object detection models,taking into account the detection speed and accuracy of the network model,and ultimately selects the YOLOv5 algorithm as the basic algorithm of this thesis.After comparing the accuracy,algorithm complexity,and model training cost of four different YOLOv5 object detection models,the YOLOv5 s model was ultimately selected as the basic algorithm for aluminum defect detection.(2)Because the size and number of detection frames in different data sets are different,in order to make the model regression better,K-Means clustering algorithm and auto anchor algorithm are used to regenerate the detection anchor frame of the model.The use of auto anchor is 12.9% higher than the m AP without the preset anchor frame.To enhance the ability of feature pyramid modules to extract and transmit feature information,three different feature pyramid modules,Sim SPPF,SPPCSPC and ASPP,were used to improve the feature pyramid structure(SPP)of YOLOv5.After the improvement,the m AP of each model increased by 2.9%,1.9% and 1.7%.In order to enhance the classification and regression capabilities of the model,the use of Decouple head to improve the detection head of YOLOv5 resulted in a 2.3% increase in model m AP.(3)Make the network model pay emphasis on useful feature information during training,SE.ECA and CBAM modules of various attention mechanisms were added to the backbone network and CSPX-1 structure respectively.Among them,after adding the attention mechanism module to the backbone network,the model m AP increased by 0.3%,1.2% and 1.0% respectively,and after adding it to the CSPX-1 structure,the model m AP increased by 1.8%,2.9% and 2.5% respectively.After adding the attention mechanism module to the CSPX-1 structure,we improved the detection header of YOLOv5 using Decouple head.The corresponding model m AP was increased by 3.2%,1.5%,and 3.7%respectively.The ECA module is not as effective as after the CSPX-1 structure,and the model m AP corresponding to SE and CBAM is well improved.(4)In order to deploy the model on mobile devices and embedded devices,lightweight networks Moblie Net V3,Shuffle Net V2 and Ghost Net were used in this thesis to replace the feature extraction network Backbone of YOLOv5.Using deep separable convolutions can effectively reduce model complexity.After replacing the backbone structure with a lightweight network,on the premise of minimal changes in the network model m AP,the number of model parameters decreased by 28.4%,45.8% and 32.2%,respectively.The amount of computation decreased by 28.8%,50.6% and 49.9%respectively.Because the Neck part of the CSPX-2 module parameter number was still very large,the replaced the lightweight backbone structure using the C3 Ghost module to replace the Neck part of the CSPX-2 module to further reduce the network parameter number and calculation amount.After improving the CSPX-2 module of the Neck part,the number of model parameters decreased by 59.6%,59.5% and 45.8%.The amount of computation decreased by 66.9%,61.9% and 62.5%.(5)In response to the black box problem of deep learning,in order to demonstrate that the network model can learn important feature information,image feature map visualization and thermal map visualization analysis were performed.The transfer of trained models to steel and PCB circuit board datasets for testing also achieved good results.