| With the rapid development of the electronic information industry,the demand of the integrated circuit industry is growing exponentially,and the maskless lithography machine based on machine vision is one of the important equipment for the development of the integrated circuit manufacturing field.The PCB layout transfer from the computer to the PCB substrate requires the maskless lithography machine to use the vision system to calibrate the relative position with the PCB substrate before completing the subsequent exposure work.In response to the current problems of visual alignment caused by unclear Mark dots and much noise in lithography,thesis focuses on researching and improving the accuracy of circular Mark dot detection in maskless lithography alignment technology,and the main research work is as follows:(1)The study is based on an improved Richer Convolutional Features(RCF)network model.First,histogram enhancement and filtering operations are applied to the image to remove the effect of noise on edge detection.In the RCF network model,fuzzy sampling is used instead of maximum pooling to accomplish the feature compression of the target image,to optimize the model parameters,and to maintain the image translation invariance property.To expand the perceptual field in the convolution process,the convolution layers of stage4 and stage5 of the RCF network model are replaced by the null convolution.Considering the different characteristics of the output images of each convolution stage,the weight proportion is added to the loss function of the RCF algorithm and the residual module is used to fuse the output images of each stage.(2)Zernike moment optimization algorithm is used to achieve sub-pixel edge detection of the image.The third-order grayscale ideal model is used in the Zernike moment algorithm to solve the problem of errors in the second-order grayscale model detection.To address the problem of difficult selection of image edge step grayscale thresholds in the Zernike moment algorithm,a particle swarm-maximum interclass variance algorithm is used for automatic selection of optimal thresholds.The improved Zernike moment algorithm can control the error of image subpixel edge point localization within 0.2pixel by simulation experiments.The detected subpixel edge points are applied to the improved random Hough circle detection,and the simulation verifies that the circle center localization error is within 0.001pixel range and the radius error is within 0.lpixel range.(3)Mark point edge detection software algorithm development based on QT and C++is completed.The improved algorithm in thesis is compared with Zernike moment and random Hough transform algorithm,and the results show that the accuracy of hole type Mark point radius detection is improved from 0.3pixel to 0.15pixel(1.65 μm),and the accuracy of inner film type Mark point radius detection is improved from 0.33pixel to 0.2pixel(2.2μm),and the stability of the algorithm is verified after several experiments,and the requirement of 30μm line width for maskless lithography exposure is satisfied. |
