An Intelligent Detection System For Intermediate Shell Surface Defects Based On Deep Metric Learning

With the rapid development of the manufacturing industry,surface defect detection of parts has become an indispensable part of the manufacturing industry.Surface defects will have a serious impact on the quality and performance of parts,even lead to the failure of parts.In recent years,with the continuous progress of artificial intelligence technology,defect detection based on machine vision has achieved a lot of development and application.However,these methods have some problems in practical application,such as insufficient generalization,low classification accuracy,slow convergence and easy to be disturbed by noise.To solve above problems,the surface defect detection method based on depth metric learning and label noise filtering method based on O2U(Overfitting to Underfitting)-Net are studied in this paper,and a set of the intelligent detection system for intermediate shell surface defects is developed by combining these two methods.The main research contents of this paper are as follows:1.In order to alleviate the influence of label noise in surface defect detection of parts,a label noise filtering method based on improved O2U-Net was proposed.Firstly,partial clean samples were used to pre-train models with the MIC(Masked Image Consistency)module,which increased the difference between difficult samples and mislabeled samples to reduce the risk of the model identifying difficult samples as mislabeled samples.Then,in order to improve the data utilization rate,the image features extracted from each batch were saved to the memory module,and the similarity between each sample was compared to further screen the label noise.Finally,the denoising performance of the proposed method was verified by the data set of intermediate shell surface defects.The experimental results show that the proposed method can achieve 97.54% accuracy for samples with different noise ratios,which can increase the accuracy by 8.77% compared with the original method.2.For the problems of low classification accuracy,slow convergence and low stability when most deep learning models are applied to surface defect detection of complex parts,a surface defect detection method based on Deep Metric Learning was proposed.Firstly,the Deep Attention Center Loss function is introduced,and the loss is weighted according to the distance from each sample to the center.The classification accuracy and convergence rate of the model are improved while the discrimination degree of background and feature importance is greatly increased.Then,the improved O2U-Net is used for sample screening and pre-training,and the centers of various samples are extracted as the initial centers of center loss,which effectively reduces the influence of label noise,improves the problem of center randomness,and reduces the probability of model training collapse.Finally,the model was tested with the data set of intermediate shell surface defects.The experimental results show that the classification accuracy of this method can reach 97.14%,which proves its superiority in the detection of parts surface defects.3.In order to realize automatic detection of surface defects of parts,a set of intelligent detection system for surface defects of intermediate shell is designed in this paper.Firstly,the types and locations of surface defects of intermediate shell are analyzed,and the overall planning of detection system is carried out according to the characteristics of defects.Then,the hardware equipment such as camera,lens and light source is selected,and the combination scheme of related industrial camera and light source is designed to complete the image acquisition work.Finally,the intelligent surface defect detection software is developed based on.NET platform.The software has functions such as straightness detection,circle detection,edge detection and Deep Learning,and integrates interfaces such as template management,data statistics and result display.Through the functional test,it is proved that the system designed in this paper can realize the high precision automatic detection function of the intermediate shell surface defects,and meet the needs of enterprises.