Machine Vision Based Typical Defects Inspection For Curved Glass Of Mobile Phone

Curved glass is more and more used in mobile phones,but it will produce some appearance defects in the production process of mobile phone curved glass.Therefore,it is necessary to detect the curved glass defects with an effective method.Machine vision technology has the advantages of high efficiency,objectivity and good stability.This paper,taked mobile phone curved screen as the research object and based on machine vision,extracts and identifies the four typical defects of scratch,pitting,chipping and burn.Firstly,the technical difficulties in the acquisition of the curved glass defect image of mobile phone are analyzed,which can be divided into the defect image acquisition method of the plane part,the curved side part and the R angle part of the mobile phone curved glass.For the plane part,the line sweeping platform is applied to obtain the image.For the curved part,an image acquisition method based on modeling analysis is proposed.For the R angle part,an image acquisition method based on multifocal image fusion is proposed.Secondly,the defect extraction algorithm of three parts of mobile phone curved glass is studied.In the plane part of curved glass,a defect extraction algorithm based on image matching is proposed.In the curved part,a defect extraction algorithm based on connected domain analysis and area threshold segmentation is proposed.For the R angle part,a multifocal Image fusion defect extraction algorithm is proposed.Then,the classification method of curved glass typical defect recognition is studied.A non-end-to-end defect classification method based on convolutional neural network is proposed.A CNN network model inclued five-convolution layer,five pooling layers and two full connections are designed,optimized by the Adam algorithm,and finally the recognition classification is completed.The experimental results indicate that under the same CNN network structure,the non-end-to-end defect identification method has a higher recognition rate and lower requirements for the CNN network model.Finally,the hardware platform construction and software framework design of the mobile phone surface glass defect detection system are carried out,and the defect extraction experiment and defect classification experiment were performed.The experimental results show that the three-part defect extraction algorithm proposed in this paper can achieve 95.3% accuracy.In the classification recognition experiment,the recognition rate based on the non-end-to-end defect classification method was 98.6%.