Research On Concave-convex Defect Detection Method Of Mobile Phone Glass Panel

With the development of information technology,mobile phone plays an increasingly important role in people's daily life.As the only output part of mobile phone image information,the quality of mobile phone screen directly affects the sales and user experience of mobile phones.In order to meet the needs of users,there are quality detection links in the production of mobile phone glass panel.As a large number of defects,bump defects play an important role in the quality detection.Now,it is difficult to meet the needs of modern enterprises by the method of artificial detection.In recent years,the rapid development of computer vision technology makes it possible to realize the automatic detection system.Therefore,this paper studies the automatic detection method of concave convex defects of mobile phone glass panel based on computer vision technology.Through the performance comparison of defect detection algorithms,and the comparative analysis of the detection results produced by different detection algorithm frameworks,combined with the task characteristics and requirements of mobile phone glass panel concave convex defect detection,the two-stage target detection Mask R-CNN framework is selected as the core framework of defect detection.In order to solve the problems of long time,low accuracy and hard to design anchor frame parameters for defect target with large aspect ratio transformation,this paper analyzes the imaging characteristics of concave convex defects in mobile phone glass panel,and designs the bright field imaging enhancement algorithm which can highlight the location information of defect to quickly and accurately locate the defect area.At the same time,this paper improves the feature extraction process in many aspects according to the characteristics of defects.Aiming at the characteristics of large scale transformation and weak semantics of defects,a backbone network based on multi-scale feature extraction is designed.At the same time,due to the large change of defect morphology,deformation convolution is integrated to extract more effective features of target defects.Aiming at some defects which are more easily distinguished by context information,a method integrating context information into ROI feature map is proposed.For the problem of dislocation between the horizontal suggestion candidate box and the defect area caused by the arbitrary orientation of the defect,a detection method based on the rotating box is designed,which not only makes the positioning more accurate,but also makes the extracted features more relevant to the target defect.Finally,the experimental verification and result analysis are carried out.In order to solve the problem that the number of mobile phone glass panels is limited and the number of different types of defects is unbalanced,this paper not only uses the commonly used transfer learning and data enhancement methods,but also designs an enhancement method specifically for the defect target,so as to increase the number of samples and balance the number of various types of defects.After finishing the fine-tuning training of the model,some feature channels and detection results are visualized,and the detection results of the proposed method are compared and analyzed in the form of evaluation index,which verifies the effectiveness of the proposed method.