Research On Moire Removal Agorithm Based On Deep Learning

The main component of electronic products,LCD screen,needs defect detection before leaving the factory,which is a crucial step in quality inspection.Defect detection requires the use of industrial cameras for image acquisition.During the acquisition process,there will be Mohr marks,which will cause great interference to the defect detection process,resulting in misjudgment between good and bad products and affecting the efficiency of defect detection.Therefore,this thesis studies the related problems of industrial screen moire,and determines that the ultimate purpose of Moire removal in this scene is to restore the original image and ensure that the defect detection system can correctly identify the quality of electronic screen.Based on this purpose,this thesis mainly contributes the following aspects:(1)Mathematical modeling of industrial screen moire based on simulated grating and Gaussian fitting function.This thesis creatively establishes the mathematical model of industrial screen moire image from two aspects:texture structure and gray distribution.The mathematical expressions of moire and grating are constructed,and the Gaussian fitting algorithm is used to fit the gray distribution of moire.The Kolmogorov Smirnov test(K-S)is used to verify that the gray distribution basically conforms to the Gaussian distribution.(2)Due to the confidentiality of industrial scenes and the scarcity of moire images on industrial screens,this thesis proposes cascaded conditional residual generative adversarial net works(CCRGAN)to generate data sets.Cascade two network model training,and improve and optimize the structure of generator and discriminator.The Wasserstein generative adversarial networks(WGAN)model,the cascaded WGAN model and the CCRGAN model in this thesis are compared.The advantages of CCRGAN model in this scene are verified by subjective evaluation,objective evaluation and image quality detection by industrial defect detection system.(3)In order to remove the moire pattern in industrial screen defect detection scene,this thesis proposes a removal algorithm based on multiscale fusion neural network(MFNN).The feature extraction part is built based on the residual block,and the idea of feature pyramid is introduced to multi-scale feature fusion training.The experimental results show that in the same defect detection algorithm,the detection rate of dark spots,dark blobs,bright spots and bright blobs is increased by 5%,5%,10%and 10%respectively.Therefore,the industrial defect moire image not only removes the moire,but also retains the original background and defect information of the image,which improves the anti moire ability of the industrial defect detection system,and is of great significance to the defect detection scene of the industrial screen.