Research On Technology Of Defect Detection Of Mobile Phone Screen Based On Deep Learning

In the industrial field,the use of optical images to detect defect on the surface of the workpiece is relatively common.The research object of this subject is the mobile phone screen.A mobile phone repair manufacturer needs to automatically detect the surface defects of the mobile phone screen.However,traditional image processing methods are difficult to adapt to the complex and diverse surface conditions of mobile phones.Therefore,this subject attempts to explore a method of detecting complex surface defects based on deep learning,and hopes to produce some inspiration for the detection of other workpieces.Based on this demand and goal,this article mainly works as follows:The specific task requirements of the subject are analyzed and the overall design of the system is formulated.This subject adopts optical images for research.A complete defect detection system includes the image acquisition module,the image preprocessing module,and the defect detection module.These three modules are finally combined in an interface platform,this article discusses these four parts,focusing on the defect detection part.This article selects the appropriate light source and camera,and builds an image acquisition system.In the image processing part,the edge-based segmentation algorithm was selected to extract the region of interest(ROI);the histogram equalization algorithm was selected for image enhancement.The defect detection part selects the semantic segmentation network,and the subject first adopts the classic U-Net network for experiments.This paper focuses on the problem of small target segmentation,that is,the problem that the proportion of defective parts is too small causes segmentation difficult,a special loss function scheme is formulated,the combination of focal-loss and tversky-loss is used,and a special initialization method is customized.Appropriate regularization methods,optimizers and evaluation standards have been established;a neural network training platform has been built.After the network training,a certain effect has been achieved.The Intersection over Union(IOU)reaches 0.669,and most of the obvious defects have been effectively detected,but system fails to detect some insignificant defects,and there are serious false detection phenomena.After analyzing the above results,the idea of first classifying and then segmenting is put forward,that is,the image is judged by the classification network to determine whether there is a defect in the image,and then the image that is determined to be defective is to be segmented.The segmentation network is added to the classification network,and the accuracy rate is greatly improved compared to the classic classification network;the Attention mechanism is added to the U-Net network,and the Attention U-Net is built to better solve the problem of small target segmentation;The above two networks are combined to design a composite network.This subject has also designed several sets of controlled experiments.According to the index results of the controlled experiments,it is known that the results of Attention U-Net are better than U-Net on the defective data set,demonstrating that the Attention mechanism can indeed effectively improve the problem of small target segmentation;the effect of the scheme of the classification network + Attention U-Net and the scheme of composite network is better than U-Net network,demonstrating the correctness of the idea of classification before segmentation.The results of the solution of classification network + Attention U-Net reaches 0.892,and the composite network reaches 0.768,both exceeding the expected target of 0.75.Finally,the interface development tool Tkinter is used to complete the construction of the interface platform,and the results are visualized to verify the detection effect.After comparing with the previous U-Net results,it is concluded that the detection effect of the scheme of classification network + Attention U-Net has indeed improved the results than U-Net network;the test of speed of the system has been tested under different hardware conditions,and the results show that the system can meet the actual requirements.