• Acta Photonica Sinica
  • Vol. 51, Issue 6, 0611002 (2022)
Jupu YANG1、2, Jialin DU1、2, Fanxing LI1, Qingrong CHEN1, Simo WANG1、2, and Wei YAN1、*
Author Affiliations
  • 1Institute of Environmental Optics,Institute of Optics and Electronics,Chinese Academy of Sciences,Chengdu 610209,China
  • 2University of Chinese Academy of Sciences,Beijing 100049,China
  • show less
    DOI: 10.3788/gzxb20225106.0611002 Cite this Article
    Jupu YANG, Jialin DU, Fanxing LI, Qingrong CHEN, Simo WANG, Wei YAN. Deep Learning Based Method for Automatic Focus Detection in Digital Lithography[J]. Acta Photonica Sinica, 2022, 51(6): 0611002 Copy Citation Text show less
    Schematic diagram of a deep learning based digital lithography autofocus system
    Fig. 1. Schematic diagram of a deep learning based digital lithography autofocus system
    Image of the centroid of each out-of-focus range on the CCD
    Fig. 2. Image of the centroid of each out-of-focus range on the CCD
    Coarse check focus network structure
    Fig. 3. Coarse check focus network structure
    Bottleneck modules
    Fig. 4. Bottleneck modules
    Confusion matrix for network models with different layer structures on the test set
    Fig. 5. Confusion matrix for network models with different layer structures on the test set
    Training results of ResNet28+FF
    Fig. 6. Training results of ResNet28+FF
    Normalized evaluation curves for different definition evaluation functions in a set of out-of-focus images
    Fig. 7. Normalized evaluation curves for different definition evaluation functions in a set of out-of-focus images
    Search algorithm flow chart
    Fig. 8. Search algorithm flow chart
    Experimental system diagram
    Fig. 9. Experimental system diagram
    Coarse focus check process display
    Fig. 10. Coarse focus check process display
    Precision check focus process display
    Fig. 11. Precision check focus process display
    Network modelAccuracy/%
    ResNet1870.0
    ResNet3477.5
    ResNet5077.5
    Table 1. Accuracy of network models with different layer structures on the test set
    Network modelAccuracy/%
    ResNet18+FF77.5
    ResNet34+FF83.1
    ResNet50+FF84.5
    Table 2. Accuracy of different layer structured network models with added feature fusion modules on the test set
    Image definition evaluation functionTime taken to evaluate a single image /s
    Laplacian0.005
    Energy0.998
    Brenner0.247
    SMD0.991
    variance0.771
    Table 3. Time required to process images with different definition evaluation functions

    Off-focus

    volume/μm

    Coarse check focus resultsSteps used for precision focus checksTotal time/msFocus check results
    -40Grade 4 negative defocus 98%6263.8True
    -29Grade 3 negative defocus 53.8%5238.4True
    -18Grade 2 negative defocus 88%4212.1True
    -10Grade 1 negative defocus 58.4%3191.4True
    -1Coarse focal plane 80%2168.9True
    9Grade 1 positive defocus 54.8%3188.8True
    17Grade 2 positive defocus 78.6%4212.9True
    28Grade 3 positive defocus 73.5%5241.1True
    39Grade 4 positive defocus 96.8%6263.5True
    Table 4. Focus detection performance of the same pattern in different out-of-focus situations using the proposed method
    Off-focus volume /μmSteps used for focus checksTotal time/msFocus check results
    -40411 023.1True
    -2930729.3True
    -1819478.1True
    -1011302.4True
    -1275.6True
    910298.1True
    1718520.5True
    2829804.1True
    39401 055.2True
    Table 5. Focus detection performance of the same pattern in different out-of-focus situations using conventional methods
    Focus check graphicsCoarse check focus resultsTime for coarse focus checks /msSteps used for precision focus checksTotal time for precision focus checks /ms
    Grade 3 positive defocus78.6%93.75148.6
    Grade 3 positive defocus73.5%93.75145.5
    Grade 3 positive defocus61.8%93.75147.2
    Table 6. Comparison of focus detection performance of different patterns at 28 μm out of focus
    Focus check graphicsCoarse check focus resultsTime for coarse focus checks/msSteps used for precision focus checksTotal time for precision focus checks/ms
    Grade 2 negative defocus68.4%93.74123.6
    Grade 2 negative defocus88%93.74123.3
    Grade 2 negative defocus76%93.74125.6
    Table 7. Comparison of focus detection performance of different patterns at -18 μm out of focus
    Focal plane detection error/μmNumber of times each error occurs in the coarse check focusNumber of times each error occurs in the precision check focus
    012
    ±1315
    ±243
    ±330
    ±420
    ±530
    ±610
    ±730
    Table 8. Results of the coarse and precise focus detection errors in 20 tests
    Jupu YANG, Jialin DU, Fanxing LI, Qingrong CHEN, Simo WANG, Wei YAN. Deep Learning Based Method for Automatic Focus Detection in Digital Lithography[J]. Acta Photonica Sinica, 2022, 51(6): 0611002
    Download Citation