Research On Surface Defect Detection Of Metal Components Based On Deep Learning

Metal components are an important class of parts in the modern automotive manufacturing industry,and precision metal components such as automotive bearings and electronic connectors produce a variety of surface defects in the production process,which are not only difficult to detect but also lead to false welding in the welding process,seriously affecting the quality of the product.At the same time,the automotive market’s demand for precision in metal components is increasing,and it is becoming increasingly important to accurately and quickly detect and identify defects on the surface of metal components in order to improve the product qualification rate and its competitiveness.However,the past methods of detecting defects by human eyes have many problems,such as high cost,low detection accuracy,slow speed,leakage and false detection,etc.These problems seriously affect the productivity of enterprises.To solve these problems,this thesis designs and develops a detection system for surface defects of metal components based on deep learning.The core of the system is to use the improved YOLOv5 detection algorithm to realize the detection of surface defects of metal components,and the main research contents are specified as follows:(1)Analyze the types of metal component surface defect detection and produce a dataset.Field research was conducted on the factory floor to understand the types of defects and the requirements for detection accuracy and speed,and images of surface defects on metal components were collected to create a dataset.To solve the problem of small number of samples and single sampling background in the original dataset,the collected images are labeled with defects and then enhanced with Mosaic and Copy-Paste methods to complete the construction and expansion of the dataset.(2)Analysis and selection of neural network models for defect detection.Four detection algorithm models of YOLOv5 with different sizes of s,m,x,and l were used for training and comparative analysis of results on the self-made dataset.The experimental results show that YOLOv5 s has a mean average precision(m AP)of83.6% and a frame rate of 25 FPS on the metal component surface defect test set.Compared with the remaining three models,YOLOv5 s has a more balanced detection accuracy and speed,which is more suitable for surface defect detection of metal components in actual production.(3)Analyze and improve the YOLOv5 s algorithm to enhance the defect detection performance.Firstly,a small number of important feature maps are generated using traditional convolution in the backbone network,and then a low-cost convolution module is used to generate a large number of redundant feature maps based on these important feature maps to reduce the computational complexity of the model training.Second,for the case of metal components with small objects and objects with large scale variations,a weighted Bidirectional Feature Pyramid Network(Bi FPN)is used to remove nodes that contribute less to feature fusion and give feature weights to make the network have the ability to distinguish the importance of features and improve the The fusion ability of multi-scale weighted features is improved.Then,to address the difficulty of learning the location information of the surface defects of metal components,Coordinate Attention(CA)is used to embed the coordinate information into the channel attention to improve the object localization of the model and enhance the perceptual field without consuming a lot of computational resources.Finally,experimental results on the metal component surface defect dataset show that the improved YOLOv5 s detection algorithm achieves a mean average precision of 95.2%and an average detection time of 33.3 ms,which improves the mean average precision by 11.6% and the average detection time by 6.7 ms compared with the original YOLOv5 s detection algorithm,effectively improving the performance of the detection algorithm.(4)Implementing a deep learning-based surface defect detection system for metal components.A Py Qt5 tool library-based metal component surface defect detection system is designed and developed using the Python programming language.The hardware part builds an image acquisition environment,and the software part implements an image detection function module,which mainly contains a detection function module,a model management module and a data storage module.After testing and verification,the improved YOLOv5 s defect detection algorithm deployed on the system runs stably and can meet the actual requirements of both accuracy and speed of metal component surface defect detection,which helps to improve the production efficiency and reputation of manufacturing enterprises.