• Optics and Precision Engineering
  • Vol. 32, Issue 10, 1538 (2024)
Heng YUAN1, Xiaoxue WANG1,*, Tinghao YAN1, and Shengchong ZHANG2
Author Affiliations
  • 1College of Software, Liaoning Technical University, Huludao2505, China
  • 2Key Laboratory of Optoelectronic Information Control and Security Technology, Tianjin300308, China
  • show less
    DOI: 10.37188/OPE.20243210.1538 Cite this Article
    Heng YUAN, Xiaoxue WANG, Tinghao YAN, Shengchong ZHANG. Cross-level feature aggregation image enhancement with dual-path hybrid attention[J]. Optics and Precision Engineering, 2024, 32(10): 1538 Copy Citation Text show less
    Network structure diagram of this paper
    Fig. 1. Network structure diagram of this paper
    Structure diagram of MDAR
    Fig. 2. Structure diagram of MDAR
    Structure diagram of DHAB
    Fig. 3. Structure diagram of DHAB
    Structure diagram of CFAM
    Fig. 4. Structure diagram of CFAM
    Visual comparison of different algorithms on the LOL dataset
    Fig. 5. Visual comparison of different algorithms on the LOL dataset
    Visual comparison of different algorithms on the MIT-Adobe 5K dataset
    Fig. 6. Visual comparison of different algorithms on the MIT-Adobe 5K dataset
    Ablation experiment visualization of different network modules
    Fig. 7. Ablation experiment visualization of different network modules
    方法PSNR/dB↑SSIM↑LPIPS↓NIQE↓
    BPDHE10.9360.4150.4397.985
    CLAHE13.4140.5510.3616.636
    LIME16.7580.5640.3958.058
    BIMEF13.8750.5750.3267.699
    RetinexNet16.7740.4250.4748.879
    KinD17.6470.7710.5145.189
    URetinex Net19.8410.8260.2354.722
    DSLR14.9820.5960.3764.416
    Uformer18.5470.7210.3214.443
    ALL-E18.2160.7630.2194.200
    EnlightenGAN17.4820.6510.3225.807
    ZeroDCE14.8600.5620.3357.767
    ZeroDCE++16.5700.5910.3285.311
    RUAS18.2260.7170.2706.340
    SCI14.7800.5220.3396.617
    Ours22.3470.8500.1784.153
    Table 1. Objective evaluation results of different algorithms on the LOL dataset
    方法PSNR/dB↑SSIM↑LPIPS↓NIQE↓
    BPDHE12.8550.5720.4415.280
    CLAHE15.9550.6080.3755.353
    LIME13.3030.7490.3194.172
    BIMEF17.9680.7970.2984.598
    RetinexNet12.5140.6700.3654.841
    KinD16.2030.7840.2544.242
    URetinex Net14.1840.7540.2423.867
    DSLR20.2430.8280.1534.352
    UFormer21.9170.8700.1853.961
    DDNet18.3080.7760.2203.830
    EnlightenGAN17.9050.8360.2383.865
    ZeroDCE15.9310.7660.2193.882
    ZeroDCE++14.6110.4050.2313.850
    RUAS15.9950.7860.1414.048
    SCI17.4770.8400.2164.122
    Ours22.7030.9030.1373.822
    Table 2. Objective evaluation results of different algorithms on the MIT-Adobe 5K dataset
    模块PSNR/dB↑SSIM↑Time/s↓
    无MDAR21.5050.8610.100
    无PMFB21.8620.8780.108
    无DHAB22.5120.8960.109
    无CFAM21.1730.7320.106
    无混合空洞卷积21.9010.8900.110
    Ours22.7030.9030.112
    Table 3. Comparison results of evaluation indicators of different network modules
    BPDHECLAHELIMEBIMEFRetinex NetKinDURetinex NetDSLR
    0.10410.11560.49140.12800.12000.05000.4000.125
    Running time↓UFormerALL-EEnlighten GANZero DCEZero DCE++RUASSCIOurs
    0.1211.0550.0570.0030.00190.0060.0010.112
    Table 4. Average enhancement time of different method
    Heng YUAN, Xiaoxue WANG, Tinghao YAN, Shengchong ZHANG. Cross-level feature aggregation image enhancement with dual-path hybrid attention[J]. Optics and Precision Engineering, 2024, 32(10): 1538
    Download Citation