Research On Mobile Phone Screen Defect Detection Based On Autoencoder

In recent years,the smartphone industry has grown rapidly,the demand for screens in the mobile phone market has continued to grow,and manufacturers have become increasingly stringent in controlling the quality of mobile phone screens.In the production process of mobile phone screens,it is inevitable that screen defects will be caused,bringing risks to the manufacturer’s reputation.Therefore,defect detection of mobile phone screens is an important part of quality control.For the traditional detection methods are not universal,the parameters are poorly adaptable,and deep learning models need a large amount of defect data for training and other problems.This thesis implements defect detection based on autoencoder,using unsupervised learning,specifically divided into two sub-tasks: defect segmentation and defect classification.In the task of defect segmentation,defect segmentation is realized by learning from normal samples,which is not restricted by defect data and does not need to mark defect label information.To improve the accuracy of defect segmentation,a defect segmentation method based on image reconstruction is proposed.A background reconstruction network is constructed through a denoising autoencoder.The background texture image is reconstructed from the defect image.The defect image through background texture suppression and threshold segmentation.Finally,the segmentation results under the multilayer features are merged to obtain the defect segmentation map.In the image reconstruction network,combined with the improved loss function,the content information and texture information of the image are learned to improve the reconstruction performance.In the segmentation process,according that the gray histogram of the residual image is a single peak,a fast and adaptive threshold segmentation method is used for defect segmentation to extract the defect part accurately.In the task of defect classification,the traditional features and the deep features are compared in classification experiments.The traditional features are selected the shape feature and color feature of the image.The deep features are extracted by a deep feature extractor obtained through sparse autoencoder.A feature fusion defect classification method is proposed.The traditional features and depth features are extracted through the fully connected layer of the multi-layer perceptron to eliminate the differences between features and improve the ability of feature expression.This thesis implements defect segmentation and defect classification through autoencoder,which provides new research ideas for practical industrial applications.On the mobile phone screen defect data set constructed in this work,the defect segmentation accuracy reaches 90.3%,which is more accurate and adaptive than other segmentation methods.The defect classification accuracy rate reaches 99.7%,which is stronger than other classification methods for feature expression ability.