Research On Deep Learning Method And Its Application In Defect Detection For Polarizer

During the manufacturing process of products,it is inevitable to produce defects.These defects will not only cause the product's performance to decline,and they may even lead to the hidden danger of safety.With the increasing demand of users and production enterprises for the quality of products,online defects detection of the product has become an important link.Using machine vision to detect defects is an effective way to realize intelligent of equipment.In actual industrial production,due to the variety kinds and shapes of the defects on products,traditional machine vision image processing methods are difficult to accurately describe defect features,and the the effectiveness of defects feature extraction is low.As a result,the accuracy of defect detection and classification is difficult to meet the demand of industrial.Deep learning is a method that allows a computer to automatically learn the features of a pattern,and integrates feature learning into the process of building a model,thus reducing the incompleteness caused by artificial design features.Therefore,this paper uses deep learning methods to detect and classify defects of polarizer,and designs a deep learning image classification network to improve the accuracy of defect classification.However,the complexity of the deep learning model easily leads to a sharp increase in the time complexity of the algorithm and the excessive cost of computing resources,combining with the real-time requirements of polarizer defect detection in industrial production,deep learning model compression networks are proposed,under the premise of ensuring the accuracy of defect classification,effectively reduces the dependence of the model on the hardware environment,improves the speed of online learning,and meets the real-time nature of online detection.The main research contents and innovative works of this paper are summarized as follows:(1)Aiming at the problems that the traditional convolutional layer of the classical deep learning network Alex Net has limited ability to extract defect features,and the fully connected layer of Alex Net has large amout of parameters,this paper proposed a deep learning classification network based on the multilayer convolutional module.The first is to replace the traditional convolutional layer with a multilayer convolutional module,which realizes the interactions of cross-channel information,enriches the extracted defect features,and improve the representation ability of the network.The second is to replace the fully connected layer with a global average pooling layer,which greatly reduces the amount of netwok parameters,and avoids overfitting.The experimental results show that the total number of parameters of the deep learning classification network based on multilayer convolutional module is one order of magnitude less than Alex Net,and the classification accuracy of polarizer is 2.4% higher than that of Alex Net.(2)In order to meet the real-time requirements of industrial applications,this paper proposes a model compression network based on parallel modules.We mainly mix different convolution template sizes to construst parallel module,which improves the ability of the module to extract more defect features.Then,the depthwise separable convolution is used to replace the traditional 3×3 convolutional layer in this module,which significantly reduces the number of parameters and the computation.The experimental results show that the classification accuracy of the model compression network constructed using parallel module is 0.5% higher than that of Mobile Net,the classification speed is nearly doubled than that of Mobile Net,the total number of parameters and multiply-accumulate operations are reduced by two orders of magnitude,and the size of the model is 95.78% smaller than Mobile Net.(3)In order to further compress the network,an efficient lightweight model compression network is proposed.First,a batch normalization layer is added after each convolutional layer of the parallel module to form a parallel normalization module,which accelerates the convergence rate of the network and avoids the problem of gradient disappearance in the process of back propagation.Second,the 3×3 depthwise convolution in the parallel normalization module is divided into 1×3 + 3×1 depthwise convolution to form a parallel asymmetric convolution module.This not only increases the ability of the network to extract defect features,but also reduces the amount of parameters and computation of the network.The experimental results show that when performing online defect detection of polarizers,the classification speed,precision,and memory consumption of the efficient lightweight model compression network constructed using the aforementioned two modules are superior to Squeeze Net and Mobile Net.(4)Aiming at the designed efficient lightweight model compression network,the optimization combination of parameters in all modules of the network is studied,and then we select an optimal set of parameters to form the final defect detection network DDN.The experimental results show that when performing online defect detection of polarizers,although the classification time per image of DDN does not change much compared to the unoptimized network,the classification accuracy is improved by 0.1% and the model size is reduced by 44.9%.(5)An automatic real-time defect detection system is developed,a human-computer interaction interface was designed,and the core components in the hardware platform are analyzed and selected.The developed system realized automatic and rapid defect detection of the image of polarizer.The interaction interface is friendly,and realizes the storage and query of defect detection results.