Research On The Defect Detection For Mobile Phone Lcd Panels Based On Fully Convolutional Networks

With the rapid development of the smart phone industry,the demand for smart phone LCD panels continues to grow.Defect detection is an important part of panel production quality control,which directly determines the overall shipment yield of mobile phone LCD panels.The existing detection technology still has a high false detection rate and a missed detection rate,and still depends on manual inspection by the naked eye.In view of this,this thesis studies the defect detection algorithm of mobile phone LCD panel based on deep learning technology,and improves the accuracy of automatic detection of mobile phone LCD panel.The thesis first analyzes the key technologies of mobile phone LCD panel defect detection,and focuses on the technology and algorithm based on fully convolutional network and two-stage convolutional neural network,and performs performance evaluation and contrast research.Combining the characteristics of the defect image of the mobile phone LCD panel and the requirements of the detection task,segmentation algorithm based on the full convolution network is selected as the core technology of the defect detection of the mobile phone LCD panel,the algorithm research and defect detection scheme design are developed.Then the defect detection task of the mobile phone LCD panel based on the full convolution network is decomposed into four steps.The first step is to construct a full convolution backbone network;the second step is to design an upsampling algorithm for the full convolution network;the third step is to fully convolve the whole convolution network.The network is trained;finally,the correction algorithm is used to correct the segmentation result.The second and fourth steps,which are the most important influences on the segmentation results,are mainly studied: the upsampling algorithm of the full convolution network and the position correction algorithm of the output segmentation map of the full convolution network.Aiming at the problem of loss of position information and low utilization of feature information due to deconvolution operation in the sampling process of the full convolutional network,the improved deconvolution operation and the structure of bidirectional feature fusion are used for the full convolution network.Upsampling operation.Aiming at the problem of positional drift of the output segmentation graph of the full convolutional network,the RPN structure is introduced to correct the output segmentation map position of the full convolution network,and the adaptive mutation genetic algorithm is used to optimize the modified parameters.Finally,the verification experiment was carried out under the dataset of the LCD panel of the mobile phone collected on the production line.The experimental results show that under the same test data set,compared with DeepLab v3,the overall segmentation accuracy mIoU of the algorithm is improved by 6.5%.