Mobile Phone Screen Defect Detection Based On Deep Learning

In recent years,with the popularization of smart phones and the rapid development of the LCD screen industry,intelligent screen detection algorithms based on image processing and computer vision have gradually replaced traditional manual detection.Mura defect in mobile phone screen,as a screen defect with various shapes and low contrast,has always been a difficult point in mobile phone screen detection.Aiming at the low degree of abstraction and poor robustness of the artificially designed Mura defect feature,this paper builds a multi-channel convolutional neural network MSFE-Net to realize the feature extraction of multi-scale defects of mobile phone screen,and adds the central loss function supervision model training.Therefore,the features of the same type of defects are internally converged,so the defect features extracted by MSFE-Net are more robust.At present,computer vision-based algorithms are mostly designed for several types of screen data,so they cannot be applied to screen data of new models.In view of the above problems,this paper proposes a self-comparison model SC-Net based on Convolutional Neural Network and Recurrent Neural Network to improve the versatility of the algorithm.The SC-Net structure includes MSFE-Net,BiLSTM,and prediction layer.The forward calculation process is as follows: First,MSFE-Net is used to extract the feature description sequence of adjacent image blocks,and then the bidirectional cyclic neural network is used to complete the context information of the feature description sequence.Contrast with global information integration,and finally use the prediction layer to get the probability value of each image block with defects.In the same mobile phone screen image,the way of comparing adjacent image blocks can offset the influence of background texture and environmental factors,thus enhancing the adaptability of the model to new data,so the versatility of the model is effectively improved.This article builds two models of mobile phone screen datasets and trains the model using only the first model of the screen dataset,while the second model's screen dataset serves as the test set.In the end,SC-Net's accuracy and recall rate on the test set is much higher than the well-known convolutional neural network AlexNet.SC-Net converts the problem of screen screen defect detection into a two-category problem,so it can only judge whether the screen has defects,and can not get the exact coordinates and defect types of the defects.In order to meet the needs of different manufacturers for the detection of mobile phone screen defects,this paper optimizes on the basis of SSD(Single Shot MultiBox Detector)model,such as using lightweight network to replace the original skeleton network,introducing focus loss,incorporating feature pyramid and using poly The class algorithm finds the best anchor parameters,which ultimately improves the accuracy and real-time of the SSD model.The experimental results show that the self-comparison model proposed in this paper has good accuracy and versatility in the open source dataset DAGM2007 and the self-built mobile phone screen dataset.At the same time,the improved SSD model can detect the accurate coordinates and categories of mobile phone screen defects in real time,which provides great help for manufacturers to improve the production proce ss of mobile phone screens.