Chinese Journal of Lasers, Volume. 51, Issue 23, 2311002(2024)

Dual-domain Multiscale Feature Merging-Based Multiresolution Temperature Tomography for Laser Absorption Spectroscopy

Jingjing Si1,4, Jingbo Wang1, Yinbo Cheng2、*, and Chang Liu3
Author Affiliations
  • 1School of Information Science and Engineering, Yanshan University, Qinhuangdao 066004, Hebei , China
  • 2Ocean College, Hebei Agricultural University, Qinhuangdao 066003, Hebei , China
  • 3School of Engineering, the University of Edinburgh, Edinburgh EH93JL, UK
  • 4Hebei Key Laboratory of Information Transmission and Signal Processing, Qinhuangdao 066004, Hebei , China
  • show less
    Figures & Tables(13)
    Schematic of multi-resolution spatial discretization model for RoI. (a) Low-resolution discretization; (b) high-resolution discretization
    Overall architecture of DMFMMnet
    Structure of MMCM
    Structure of AMFMM
    Structure of FEM
    Laser beam arrangement of experimental TDLAS sensor
    Comparison of low-resolution temperature distribution images reconstructed by three networks. (a) Sample image of low-resolution temperature distribution; low-resolution temperature distribution images reconstructed by (b) H-CNN-LR, (c) HVTMFnet-LR, and (d) DMFMMnet
    Comparison of high-resolution temperature distribution images reconstructed by six networks. (a) Sample image of high-resolution temperature distribution; high-resolution temperature distribution images reconstructed by (b) H-CNN-HR, (c) HVTMFnet-HR, (d) LBlock+Bicubic, (e) LBlock+SRCNN, (f) LBlock+ESRT, and (g) DMFMMnet
    Temperature distribution images reconstructed using real measurement data of TDLAS sensor. (a) Top view of measurement field of TDLAS sensor; low-resolution temperature distribution images reconstructed by (b) DMFMMnet, (c) HVTMFnet-LR, and (d) H-CNN-LR; high-resolution temperature distribution images reconstructed by (e) DMFMMnet, (f) HVTMFnet-HR, (g) H-CNN-HR, (h) LBlock+Bicubic, (i) LBlock+SRCNN, and (j) LBlock+ESRT
    • Table 1. Parameter settings for each module of DMFMMnet

      View table

      Table 1. Parameter settings for each module of DMFMMnet

      ModuleLayerInput dimStrideOutput dimLayerInput dimStrideOutput dim
      LBlock2×2 Conv 12×4×8(1, 1)8×3×7FC1520520
      2×2 Conv 28×3×7(1, 1)16×2×6FC25201024
      2×2 Conv 316×4×8(1, 1)32×1×5FC310241024
      3×3 Conv 12×4×8(1, 1)16×4×8FC410241600
      3×3 Conv 28×3×7(1, 1)32×3×7
      1×1 Conv 132×1×532×1×5
      IMBlock3×3 Conv 31×40×40(1, 1)32×40×401×1 Conv 216×80×1(1, 1)10×80×1
      3×3 Conv 432×40×40(1, 1)32×40×401×1 Conv 310×1×40(1, 1)16×1×40
      3×3 Conv 532×40×40(1, 1)16×40×401×1 Conv 410×1×40(1, 1)16×1×40
      3×3 Conv 616×40×40(1, 1)16×40×401×1 Conv 510×40×1(1, 1)16×40×1
      3×3 Conv 716×40×40(1, 1)32×40×401×1 Conv 610×40×1(1, 1)16×40×1
      5×5 Conv 116×40×40(1, 1)16×40×40
      HBlock1×1 Conv 732×40×4064×40×403×3 Conv 1264×40×40(1, 1)48×40×40
      1×1 Conv 880×40×4064×40×403×3 Conv 1348×40×40(1, 1)80×40×40
      3×3 Conv 864×40×40(1, 1)48×40×40FC564512
      3×3 Conv 948×40×40(1, 1)32×40×40FC65121600
      3×3 Conv 1032×40×40(1, 1)64×40×40Deconv64×40×40(2, 2)1×80×80
      3×3 Conv 1148×40×40(1, 1)64×40×40
    • Table 2. Comparison of mean PSNR and mean SSIM of low-resolution temperature distribution images reconstructed by three networks

      View table

      Table 2. Comparison of mean PSNR and mean SSIM of low-resolution temperature distribution images reconstructed by three networks

      NetworkSSIM/PSNR
      SSNR=25 dBSSNR=30 dBSSNR=35 dBSSNR=40 dBSSNR=45 dB
      H-CNN-LR0.9584/23.61180.9721/25.18650.9787/26.04550.9778/27.21390.9832/27.4768
      HVTMFnet-LR0.9870/29.31020.9904/31.95700.9919/33.76680.9951/34.85740.9957/35.3281
      DMFMMnet0.9891/29.50480.9923/32.86660.9939/34.58720.9952/36.00220.9955/36.0301
    • Table 3. Comparison of mean PSNR and mean SSIM of high-resolution temperature distribution images reconstructed by six networks

      View table

      Table 3. Comparison of mean PSNR and mean SSIM of high-resolution temperature distribution images reconstructed by six networks

      NetworkSSIM/PSNR
      SSNR=25 dBSSNR=30 dBSSNR=35 dBSSNR=40 dBSSNR=45 dB
      H-CNN-HR0.9677/24.13780.9760/25.49090.9810/26.68380.9842/27.34110.9854/27.9248
      HVTMFnet-HR0.9888/30.25840.9924/33.03550.9948/34.49210.9951/36.11580.9959/36.4694
      LBlock+Bicubic0.9879/29.49080.9931/32.76550.9944/34.56700.9951/35.42270.9943/35.4467
      LBlock+SRCNN0.9870/29.64520.9913/32.71460.9953/35.06410.9968/36.11040.9935/36.3672
      LBlock+ESRT0.9885/30.91440.9906/33.47300.9963/36.08160.9947/37.05130.9979/38.6104
      DMFMMnet0.9892/30.48840.9940/33.80990.9975/36.50900.9972/37.27510.9985/38.7245
    • Table 4. Comparison of training time and reconstruction time of 6 networks

      View table

      Table 4. Comparison of training time and reconstruction time of 6 networks

      NetworkTraining time /sLow-resolution temperature distribution reconstruction time /(×10-5 s)High-resolution temperature distribution reconstruction time /(×10-5 s)
      H-CNN-MR1532.0362.774
      HVTMFnet-MR4434.7335.357
      LBlock+Bicubic3424.593
      LBlock+SRCNN5408.863
      LBlock+ESRT421761.673
      DMFMMnet7605.00718.288
    Tools

    Get Citation

    Copy Citation Text

    Jingjing Si, Jingbo Wang, Yinbo Cheng, Chang Liu. Dual-domain Multiscale Feature Merging-Based Multiresolution Temperature Tomography for Laser Absorption Spectroscopy[J]. Chinese Journal of Lasers, 2024, 51(23): 2311002

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: spectroscopy

    Received: Mar. 14, 2024

    Accepted: May. 27, 2024

    Published Online: Dec. 11, 2024

    The Author Email: Yinbo Cheng (cyb@hebau.edu.cn)

    DOI:10.3788/CJL240684

    CSTR:32183.14.CJL240684

    Topics