Automotive Glass Defect Detection Based On YOLO Algorithm

With the increase in car production in recent years,the demand for curved laminated glass has been on the rise.As a major component of automotive accessories,curved laminated glass can play an important role in protecting the personal safety of vehicle owners.Due to its unique design,curved laminated glass is prone to bubbles,scratches,chipped edges,and other defects during the production process,affecting the product’s quality and safety.The current detection of automotive glass defects is mainly based on traditional machine vision methods,which require professionals to design glass defect image processing algorithms according to the actual situation and are highly dependent on the level of expertise of the developer.Artificial intelligence detection technology,represented by deep learning,has the advantages of efficient image feature extraction,robustness,and versatility,and is widely used in research and engineering practice.Therefore,the application of deep learning technology to the field of automotive glass defect detection is of great significance to solve the problems of complex programming and time-consuming traditional vision algorithms.In this paper,the defect detection of curved laminated glass for automobiles is studied based on the YOLO algorithm,and the main work is as follows:1.For the characteristics of defect imaging such as bubbles,stains,scratches,and chipped edges,two kinds of lighting methods,coaxial light,and backlight,are used to improve the contrast between the glass surface defects and the background and to improve the quality of the data set.2.Using offline data enhancement methods such as flip and pan,and drawing on the idea of pseudo-label generation methods,the weights obtained from small sample training are used to automatically annotate new samples;this not only achieves the expansion and diversification of the dataset and solves the shortage of glass defect datasets,but also improves the annotation efficiency of defect samples.3.The ASFF adaptive feature fusion module and CBAM attention mechanism are added to the YOLOv5 s model,and the adaptive feature fusion network YOLOv5-CA model is proposed,which improves the model’s ability to extract glass defect features and enhances the model’s detection effect on small target glass defects.The m AP of the improved model reached 91.6%,the detection speed reached 24.5ms/sheet,and the detection accuracy of the improved model was improved to a large extent.4.To improve the detection speed of the model,the model backbone network was replaced with Shuffle Netv2 lightweight network based on the YOLOv5-CA model,and an optimizer adapted to it was selected so that the model was optimized.Compared with the YOLOv5-CA model,the detection accuracy is improved by 4.3% and the detection time of a single image is reduced by 6.9%,significantly improving the model’s overall performance.5.Finally,for curved automotive glass surface defect detection,an online system is built based on the optimized YOLOv5 s algorithm for curved automotive glass surface defect detection.After testing,it is proved that the average correct detection rate of the glass defect detection software in this paper reaches 94.3%,which provides a feasible detection solution for curved automotive glass defect detection and has important theoretical and practical reference value.