Research On U-Net-based Algorithm For Mobile Phone Screen Surface Defect Detection

The detection of surface defects on mobile phone screens is an essential part of mobile phone production,which is currently done mainly by manual inspection or conventional image processing algorithms.The low efficiency of manual inspection and the low accuracy and weak robustness of conventional image processing algorithms make it urgent to develop new surface defect detection algorithms for mobile phone screens.In view of the high accuracy and strong robustness of deep learning algorithms,this paper investigates the deep learning-based surface defect detection algorithm for mobile phone screens.In order to improve the efficiency of deep learning based defect detection algorithms,this paper proposes an MB encoder-decoder semantic segmentation framework based on the MBConv module.The main component module of the framework is the MBConv module,which makes the framework very efficient.In addition,by using the same type of MBConv module in the corresponding position of encoder and decoder,it makes the decoder and encoder more compatible with each other,thus increasing the efficiency without decreasing the accuracy.To verify the superiority of the MB encoder-decoder framework,this thesis applies the MB encoder-decoder framework to U-Net networks and proposes the EU-Net network（Efficient U-Net）.Comparative experiments demonstrate that the efficiency of the EU-Net network is more than 170 times that of the U-Net network,the number of parameters is less than half that of the U-Net network,and the accuracy is improved by 0.9%.In addition,the ablation experiments also demonstrate the superiority of the MB encoder-decoder framework.Finally,the EU-Net network is applied to the surface defect detection of mobile phone screens in industrial production for the first time,which achieves the need of real-time and accuracy,and contributes to the application of deep learning to industrial defect detection.