• Opto-Electronic Engineering
  • Vol. 50, Issue 12, 230225-1 (2023)
Shengjun Xu1,2, Hua Yang1,2,*, Minghai Li1, Guanghui Liu1,2..., Yuebo Meng1,2 and Jiuqiang Han1,2|Show fewer author(s)
Author Affiliations
  • 1College of Information and Control Engineering, Xi′an University of Architecture and Technology, Xi′an, Shaanxi 710055, China
  • 2Xi′an Key Laboratory of Building Manufacturing Intelligent & Automation Technology, Xi′an, Shaanxi 710055, China
  • show less
    DOI: 10.12086/oee.2023.230225 Cite this Article
    Shengjun Xu, Hua Yang, Minghai Li, Guanghui Liu, Yuebo Meng, Jiuqiang Han. Low-light image enhancement based on dual-frequency domain feature aggregation[J]. Opto-Electronic Engineering, 2023, 50(12): 230225-1 Copy Citation Text show less
    DF-DFANet network structure
    Fig. 1. DF-DFANet network structure
    Structure of spectral illumination estimation module
    Fig. 2. Structure of spectral illumination estimation module
    Structure of multiple spectral attention module
    Fig. 3. Structure of multiple spectral attention module
    Structure of frequency domain feature aggregation module
    Fig. 4. Structure of frequency domain feature aggregation module
    LOL dataset enhancement results comparison
    Fig. 5. LOL dataset enhancement results comparison
    Comparison of enhancement results of mit-adobe fivek dataset
    Fig. 6. Comparison of enhancement results of mit-adobe fivek dataset
    Comparison of experimental effects of modular attention structure
    Fig. 7. Comparison of experimental effects of modular attention structure
    Comparison of PSNR results for module attention structure
    Fig. 8. Comparison of PSNR results for module attention structure
    Comparison of effect diagrams of modular ablation experiments
    Fig. 9. Comparison of effect diagrams of modular ablation experiments
    Test results of monitoring images of low-light vehicles at night
    Fig. 10. Test results of monitoring images of low-light vehicles at night
    MethodPSNRSSIMLPIPS
    RetinexNet[26]16.77400.42500.4739
    Zero-DCE[27]14.86070.56240.3352
    DSLR[28]14.98220.59640.3757
    KinD[29]17.64760.77150.1750
    EnGAN[30]17.48290.65150.3223
    GLAD[32]19.71820.68200.3994
    RUAS[31]16.40470.50340.2078
    R2RNet[10]20.20700.8160-
    UHDFour[8]23.09260.8720-
    URetinexNet[6]21.32820.8348-
    Ours24.37140.89370.1525
    Table 1. LOL real-world dataset results
    MethodPSNRSSIMLPIPS
    Exposure[33]18.74120.81590.1674
    CycleGAN[34]19.38230.78520.1636
    RetinexNet[26]12.51460.67080.2535
    DSLR[28]20.24350.82890.1526
    KinD[29]16.20320.78410.1498
    EnGAN[30]17.90500.83610.1425
    Zero-DCE[27]15.93120.76680.1647
    Zero-DCE++[35]14.61110.40550.2309
    RUAS[31]15.99530.78630.1397
    Ours22.72140.87260.1153
    Table 2. MIT-Adobe FiveK dataset results
    MethodPSNRSSIMLPIPS
    Baseline22.70520.81470.2078
    With serial of CA & SA23.60420.82830.1837
    With parallel of CA & SA24.37140.89370.1525
    Table 3. Comparison results of module attention structure testing
    ModelFDIEMMSAMDDFAMPSNRSSIM
    Baseline×××20.86200.8515
    Model-1×21.35820.8653
    Model-2×22.04010.8878
    Model-3×21.90680.8919
    Ours24.37140.8937
    Table 4. Experimental results of network module ablation
    ModelTime/msParams/MFLOPs/GPSNRSSIM
    RetinexNet[26]209.2136.015116.77400.4250
    Zero-DCE[27]20.975.211214.86710.5624
    KinD[29]103529.130320.37920.7715
    EnGAN[30]203361.010217.48280.6515
    GLAD[32]2511252.141019.71820.6820
    MBLLEN[36]801.9519.956017.85830.7247
    LPNet[37]180.150.770021.46120.8020
    URetinexNet[6]2.930.341801.411021.32820.8348
    Ours481.61288.377624.37140.8937
    Table 5. Comparison of different network average processing time, model size and floating-point operations
    Shengjun Xu, Hua Yang, Minghai Li, Guanghui Liu, Yuebo Meng, Jiuqiang Han. Low-light image enhancement based on dual-frequency domain feature aggregation[J]. Opto-Electronic Engineering, 2023, 50(12): 230225-1
    Download Citation