• Laser & Optoelectronics Progress
  • Vol. 60, Issue 12, 1210017 (2023)
Hongchun Yuan, Lingdong Kong*, Shanshan Zhang, Kai Gao, and Yurui Yang
Author Affiliations
  • School of Information, Shanghai Ocean University, Shanghai 201306, China
  • show less
    DOI: 10.3788/LOP221324 Cite this Article Set citation alerts
    Hongchun Yuan, Lingdong Kong, Shanshan Zhang, Kai Gao, Yurui Yang. Super-Resolution Reconstruction Algorithm of Underwater Image Based on Information Distillation Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1210017 Copy Citation Text show less
    Schematic diagram of residual network structure based on information distillation mechanism
    Fig. 1. Schematic diagram of residual network structure based on information distillation mechanism
    Residual feature distillation block
    Fig. 2. Residual feature distillation block
    Spatial attention module
    Fig. 3. Spatial attention module
    Training loss
    Fig. 4. Training loss
    Results of ablation experiments on GFF and SA (validation set)
    Fig. 5. Results of ablation experiments on GFF and SA (validation set)
    Visualization of SA
    Fig. 6. Visualization of SA
    Distilled feature maps from different stages (part)
    Fig. 7. Distilled feature maps from different stages (part)
    Comparison of visual results (RPSNR and MSSIM) of im_xb_660_ from USR-248 dataset among different algorithms (×2)
    Fig. 8. Comparison of visual results (RPSNR and MSSIM) of im_xb_660_ from USR-248 dataset among different algorithms (×2)
    Comparison of visual results (RPSNR and MSSIM) of im_xb_5341_ from USR-248 dataset among different algorithms (×4)
    Fig. 9. Comparison of visual results (RPSNR and MSSIM) of im_xb_5341_ from USR-248 dataset among different algorithms (×4)
    Comparison of visual results (RPSNR and MSSIM) of im_xb_331_ from USR-248 dataset among different algorithms (×8)
    Fig. 10. Comparison of visual results (RPSNR and MSSIM) of im_xb_331_ from USR-248 dataset among different algorithms (×8)
    SR result generated with SRIDM in real underwater scenes
    Fig. 11. SR result generated with SRIDM in real underwater scenes
    Comparison of parameters, floating-operations, and running time of different algorithms
    Fig. 12. Comparison of parameters, floating-operations, and running time of different algorithms
    ScaleBaseGFFSARPSNR /dBMSSIM
    ×4××27.56290.7142
    ×27.56600.7143
    ×27.56900.7146
    27.58510.7150
    Table 1. Results of ablation experiments on GFF and SA (test set)
    dRPSNR /dBMSSIM
    0.2527.76400.7640
    0.527.75660.7635
    0.7527.74800.7631
    Table 2. [in Chinese]
    AlgorithmScaleParams /103FLOPs /109RPSNR /dBMSSIM
    Bicubic×232.54890.9083
    SRCNN×283.132.73910.9097
    DSRCNN×22972896.033.58310.9186
    EDSR×2107492.033.62450.9189
    RCAN×21192101.633.62070.9191
    CARN×296492.733.67210.9193
    IMDN×269462.233.68100.9194
    SRIDM×266860.133.69830.9198
    Bicubic×426.77020.7432
    SRCNN×483.127.15370.7402
    DSRCNN×42972896.027.65120.7572
    EDSR×4122236.027.70670.7628
    RCAN×4134038.427.70030.7626
    CARN×4111236.227.74260.7634
    IMDN×471516.027.73360.7637
    SRIDM×468915.527.76400.7640
    Bicubic×823.57570.6080
    SRCNN×883.124.10150.6043
    DSRCNN×82972896.024.41170.6211
    EDSR×8137022.124.48360.6304
    RCAN×8148722.724.48140.6295
    CARN×8126022.124.50040.6310
    IMDN×87984.424.49070.6303
    SRIDM×87724.324.50180.6304
    Table 3. Comparison of average RPSNR and MSSIM by different super-resolution algorithms on test datasets
    Model×2×4×8
    Params /103668689772
    FLOPs /10960.1515.484.31
    Run-time /ms3.853.733.59
    Frames per second259.74268.10278.55
    Model_size /MB2.62.73.0
    Table 4. Number of params, FLOPs, run-time, and memory requirement of SRIDM on RTX 3090
    Hongchun Yuan, Lingdong Kong, Shanshan Zhang, Kai Gao, Yurui Yang. Super-Resolution Reconstruction Algorithm of Underwater Image Based on Information Distillation Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1210017
    Download Citation