Infrared and Laser Engineering, Volume. 51, Issue 12, 20220125(2022)

A review of deep learning fusion methods for infrared and visible images

Lin Li... Hongmei Wang and Chenkai Li |Show fewer author(s)
Author Affiliations
  • School of Astronautics, Northwestern Polytechnical University, Xi’an 710072, China
  • show less
    Figures & Tables(19)
    Infrared and visible image fusion framework based on convolutional neural network[43]
    Infrared and visible image fusion framework based on autoencoder network[54]
    Infrared and visible image fusion framework based on generative countermeasure network[10]
    Infrared and visible images collected by Athena, DHV, FEL and TRICLOBS systems respectively. (a1)-(d1) Infrared images; (a2)-(d2) Visible images
    The Images in roads, vehicles, and pedestrians scenes respectively. (a1)-(c1) Infrared images; (a2)-(c2) Visible images
    The images in BackyardRunner, CoatDeposit, GroupFight, and MulitpleDeposit scenes respectively. (a1)-(d1) Infrared images; (a2)-(d2) Visible images
    The first frames in two video sequences respectively. (a1), (b1) Infrared images; (a2), (b2) Visible image
    Qualitative fusion results. (a), (b) Infrared and visible images; (c)-(l) Fusion methods of DenseFuse, FusionDN, U2 Fusion, FusionGAN, DDcGAN, GANMcC, RFN-Nest, STDFusionNet, SDDGAN and SeAFusion
    • Table 1. Limitations of typical CNN-based fusion methods

      View table
      View in Article

      Table 1. Limitations of typical CNN-based fusion methods

      ReferencesLimitation
      [40] Being suitable for mutil-focus image fusion, only the last convolutional layer features are used to calculate the fusion result
      [46] The information in the middle layer is lost, and the fusion strategy has no theoretical support
      [50] The structure is simple and prone to overfitting
      [54] The model mainly saves detailed texture information and cannot highlight infrared targets
    • Table 2. Limitations of typical autoencoder-based fusion methods

      View table
      View in Article

      Table 2. Limitations of typical autoencoder-based fusion methods

      ReferencesLimitation
      [57] The model is not targeted enough to highlight the infrared target, and the fusion strategy is simple
      [64] Insufficient attention to texture information, large amount of network parameters are not conducive to application
      [68] Network channels share weights, pre-training models focus on common information, and unique information may be lost
      [67] Abundant texture details cannot be obtained
    • Table 3. Limitations of typical GAN-based fusion methods

      View table
      View in Article

      Table 3. Limitations of typical GAN-based fusion methods

      ReferencesLimitation
      [10] Insufficient consideration of infrared brightness information.
      [79] The two adversarial losses are difficult to balance, and fusion image target is distorted
      [86] Under the two-discriminator condition, the Wasserstein distance adversarial loss does not enhance the target brightness
      [89] The lack of well-segmented datasets, the quality of the pre-training model depends on the accuracy of semantic segmentation
    • Table 4. Summary of infrared and visible image fusion methods based on deep learning

      View table
      View in Article

      Table 4. Summary of infrared and visible image fusion methods based on deep learning

      TypeTypical methodsCharacteristic
      Input methodSingle channel[10], [45], [52], [54-55], [79], [82], [83] Cascading source images, mining the fusion ability of the network
      Multi-channel[46], [47], [50-51], [53], [56-72] Distinguishing the source images, but need to design a fusion strategy
      Multi-image multi-channel[67], [88-89] Inputting the source images in proportion, keeping the same category information of the source images
      Preprocess image[70-71], [86], [89-90] Providing more useful information for fused images
      Common blockAttention network[45], [51], [53], [63], [65], [85], [87] Enhancing feature maps from channels and spaces, it can be embedded in any network
      Nest network[63-65] The network structure is complex, and focusing on the shallow and middle layers of the network
      Skip connection[59], [68], [77], [87] Based on residual and dense networks,it prevents loss of useful shallow information
      Loss Func-tion Perceptual loss[55], [66], [82], [87] Balancing feature error between reconstructed image and input
      TV loss[47], [79] constraining the fused image to exhibit similar gradient variation with the visible image
      Edge detail loss[69], [82], [83-84] Enhancing fusion image edge detail
      Sematic loss[72] More targeted to different information of the scene
    • Table 5. Objective evaluation indicators of different methods in the Kaptein_1654 scene

      View table
      View in Article

      Table 5. Objective evaluation indicators of different methods in the Kaptein_1654 scene

      MethodsENMISSIMSDAGSFPSNRVIFFCCQAB/F
      DenseFuse6.4212.830.7229.743.626.9316.390.340.530.36
      FusionDN7.1914.370.6446.486.6712.9714.560.550.520.42
      U2 Fusion6.5813.160.7028.684.618.6216.170.350.530.40
      FusionGAN5.7411.470.6717.103.296.2817.050.080.640.17
      DDcGAN6.9613.930.5937.176.2911.6315.150.320.520.38
      GANMcC6.0612.110.6925.362.134.4415.380.210.560.14
      RFN-Nest6.5413.090.6831.472.394.9915.690.320.520.28
      STDFusionNet6.7013.410.6552.905.2911.2215.170.400.510.54
      SSDGAN5.8511.700.6423.061.493.6613.150.190.560.08
      SeAFusion6.7113.430.6741.075.8611.2513.920.410.560.49
    • Table 6. Objective evaluation indicators of different methods in the Sandpath scenario

      View table
      View in Article

      Table 6. Objective evaluation indicators of different methods in the Sandpath scenario

      MethodsENMISSIMSDAGSFPSNRVIFFCCQAB/F
      DenseFuse6.6813.350.6929.066.6110.8919.860.540.680.36
      FusionDN7.4214.840.5545.1310.9518.3515.750.820.680.29
      U2 Fusion6.3412.670.6920.466.4210.4919.190.360.690.33
      FusionGAN6.4312.850.6021.126.2210.3017.050.130.660.29
      DDcGAN7.2514.500.4937.8510.2316.9115.030.440.690.37
      GANMcC6.3312.650.6921.123.405.6619.440.220.710.19
      RFN-Nest6.8913.790.6432.014.888.1319.240.500.680.42
      STDFusionNet6.8213.640.5935.096.8311.6418.590.220.600.56
      SSDGAN5.9511.910.6216.262.083.5616.320.180.710.09
      SeAFusion6.8213.640.6533.027.4312.2617.810.350.660.42
    • Table 7. Objective evaluation indicators of different methods in the campus_1 scenario

      View table
      View in Article

      Table 7. Objective evaluation indicators of different methods in the campus_1 scenario

      MethodsENMISSIMSDAGSFPSNRVIFFCCQAB/F
      DenseFuse7.1314.260.6437.507.3217.0714.990.290.870.44
      FusionDN7.5615.120.6050.7911.3225.3614.550.330.860.42
      U2 Fusion7.1614.320.6237.799.0419.9414.950.300.880.40
      FusionGAN6.1512.290.5518.655.5613.0312.980.081.080.14
      DDcGAN7.3814.760.5344.1611.4824.2314.190.230.830.38
      GANMcC7.1614.330.5936.896.2111.0215.490.210.880.21
      RFN-Nest7.1914.390.5839.475.0411.1914.830.270.870.22
      STDFusionNet7.3914.790.6849.1311.3028.5215.670.170.850.50
      SSDGAN6.7013.400.5230.073.468.1113.560.180.950.12
      SeAFusion7.5615.110.5953.3611.8227.8714.210.290.860.47
    • Table 8. Objective evaluation indicators of different methods in the campus_2 scenario

      View table
      View in Article

      Table 8. Objective evaluation indicators of different methods in the campus_2 scenario

      MethodsENMISSIMSDAGSFPSNRVIFFCCQAB/F
      DenseFuse7.4114.810.6250.768.7620.7115.130.500.930.44
      FusionDN7.4414.880.6051.9910.3323.6214.850.470.910.46
      U2 Fusion7.3214.640.6052.4910.5723.9915.050.520.920.50
      FusionGAN6.7013.400.4927.516.0214.0212.230.230.970.16
      DDcGAN7.3614.720.5246.7410.1223.0914.160.280.900.37
      GANMcC7.4214.840.5548.776.2113.8815.530.450.930.26
      RFN-Nest7.4214.840.5549.856.0713.7114.800.450.940.26
      STDFusionNet7.3414.680.5854.2711.3328.3513.990.320.880.49
      SSDGAN7.0214.030.4642.714.8911.7813.700.350.930.17
      SeAFusion7.7115.420.6066.0013.7932.0013.500.480.560.92
    • Table 9. Objective evaluation indicators of different methods in the MulitpleDeposit scenario

      View table
      View in Article

      Table 9. Objective evaluation indicators of different methods in the MulitpleDeposit scenario

      MethodsENMISSIMSDAGSFPSNRVIFFCCQAB/F
      DenseFuse7.6615.310.7871.026.6515.8616.640.720.960.55
      FusionDN7.4614.910.7553.447.2216.8716.610.590.950.52
      U2 Fusion7.2314.640.7752.4910.5723.9915.050.520.970.50
      FusionGAN7.1314.270.7043.065.9814.2115.880.280.970.38
      DDcGAN7.2914.580.6747.606.8515.4715.740.360.950.43
      GANMcC7.7115.420.7573.354.4410.3815.270.570.960.36
      RFN-Nest7.7015.390.7572.454.9611.8616.140.670.990.47
      STDFusionNet7.5015.010.7272.258.8623.6619.840.690.920.62
      SSDGAN7.6715.350.7167.963.778.5815.900.510.960.25
      SeAFusion7.7915.590.7476.368.8021.4617.970.800.950.62
    • Table 10. Objective evaluation indicators of different methods in the VisitorParking scenario

      View table
      View in Article

      Table 10. Objective evaluation indicators of different methods in the VisitorParking scenario

      MethodsENMISSIMSDAGSFPSNRVIFFCCQAB/F
      DenseFuse6.7713.540.7434.644.4110.9718.180.510.670.42
      FusionDN7.5015.000.6352.267.9219.6214.320.700.630.40
      U2 Fusion6.5313.050.7431.374.5511.3019.060.410.660.40
      FusionGAN6.1912.390.6532.063.9010.0014.590.210.630.27
      DDcGAN7.2614.520.6142.807.0216.9914.840.550.660.39
      GANMcC6.7613.530.7239.092.766.6018.880.370.650.22
      RFN-Nest7.1814.360.6844.543.579.4115.220.610.660.40
      STDFusionNet6.2812.550.6826.094.8314.4320.300.190.610.49
      SSDGAN6.4912.970.7229.542.095.0119.270.340.670.13
      SeAFusion6.7913.580.7035.865.9614.8416.460.470.660.47
    • Table 11. Running time of different methods

      View table
      View in Article

      Table 11. Running time of different methods

      MethodsRun time/s
      DenseFuse3.7234
      FusionDN2.5158
      U2 Fusion1.0212
      FusionGAN0.5221
      DDcGAN2.4545
      GANMcC1.0142
      RFN-Nest1.1682
      STDFusionNet0.0480
      SDDGAN0.1970
      SeAFusion0.1605
    Tools

    Get Citation

    Copy Citation Text

    Lin Li, Hongmei Wang, Chenkai Li. A review of deep learning fusion methods for infrared and visible images[J]. Infrared and Laser Engineering, 2022, 51(12): 20220125

    Download Citation

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

    Category: Image processing

    Received: Feb. 23, 2022

    Accepted: --

    Published Online: Jan. 10, 2023

    The Author Email:

    DOI:10.3788/IRLA20220125

    Topics