Laser & Optoelectronics Progress, Volume. 59, Issue 24, 2410006(2022)

Cross-Modal Hash Method Based on Multi-Scale Fusion and Projection Matching Constraint

Wanyu Deng, Yina Zhao*, Wanzhen Yang, Bo Zhang, Hao Li, and Shuqi Ye
Author Affiliations
  • School of Computer Science & Technology, Xi'an University of Posts & Telecommunications, Xi'an 710121, Shaanxi , China
  • show less
    Figures & Tables(15)
    Framework of proposed MFPMC
    P-R curves of Image2Text on MIRFlickr-25K dataset when Hash code length is 16 bit
    P-R curves of Text2Image on MIRFlickr-25K dataset when Hash code length is 16 bit
    P-R curves of Image2Text on NUS-WIDE dataset when Hash code length is 16 bit
    P-R curves of Text2Image on NUS-WIDE dataset when Hash code length is 16 bit
    Influence of hyper-parameter ξ on mAP on MIRFlickr-25K dataset
    Influence of hyper-parameter τ on mAP on MIRFlickr-25K dataset
    Influence of hyper-parameter γ on mAP on MIRFlickr-25K dataset
    Influence of hyper-parameter η on mAP on MIRFlickr-25K dataset
    Influence of hyper-parameter μ on mAP on MIRFlickr-25K dataset
    • Table 1. Detailed parameter settings for IMFM

      View table

      Table 1. Detailed parameter settings for IMFM

      InputLayerKernel sizeStrideOutput
      Original imageAverage pooling 15×55×5Ipool 1
      Ipool 11×1Conv1×11×1IMs-feture 1
      Original imageAverage pooling 210×1010×10Ipool 2
      Ipool 21×1Conv1×11×1IMs-feature 2
      Original imageAverage pooling 315×1515×15Ipool 3
      Ipool 31×1Conv1×11×1IMs-feature 3
    • Table 2. Detailed parameter settings for TMFM

      View table

      Table 2. Detailed parameter settings for TMFM

      InputLayerKernel sizeStrideOutput
      Bow vectorAverage pooling 11×501×50Tpool 1
      Tpool 11×1Conv1×11×1TMs-feture 1
      Bow vectorAverage pooling 21×301×30Tpool 2
      Tpool 21×1Conv1×11×1TMs-feature 2
      Bow vectorAverage pooling 31×151×15Tpool 3
      Tpool 31×1Conv1×11×1TMs-feature 3
      Bow vectorAverage pooling 41×101×10Tpool 4
      Tpool 41×1Conv1×11×1TMs-feature 4
      Bow vectorAverage pooling 51×51×5Tpool 5
      Tpool 51×1Conv1×11×1TMs-feature 5
    • Table 3. Comparison of mAP values of different methods on MIRFlickr-25K dataset

      View table

      Table 3. Comparison of mAP values of different methods on MIRFlickr-25K dataset

      MethodImage2TextText2Image
      16 bit32 bit64 bit16 bit32 bit64 bit
      SCM0.61570.62130.62680.61020.62840.6292
      SePH0.64810.64530.65960.64570.64760.6508
      STMH0.58770.59010.60010.58630.58770.5879
      CMFH0.57800.58270.58610.57840.58780.5889
      DCMH0.72190.73320.74500.75260.75760.7704
      PRDH0.70520.71250.72080.76070.77390.7784
      CMHH0.73020.73870.74440.73200.72830.7301
      MFPMC0.75010.76080.76870.77640.78950.7898
    • Table 4. Comparison of mAP values of different methods on NUS-WIDE dataset

      View table

      Table 4. Comparison of mAP values of different methods on NUS-WIDE dataset

      MethodImage2TextText2Image
      16 bit32 bit64 bit16 bit32 bit64 bit
      SCM0.49050.49460.49950.45980.46600.4701
      SePH0.53240.53500.55290.50780.50950.5177
      STMH0.43540.44710.45440.38950.40980.4187
      CMFH0.39250.39580.39900.39560.39550.3978
      DCMH0.52570.53750.54580.57920.58750.5944
      PRDH0.59190.60580.61160.61550.62870.6349
      CMHH0.55300.56970.55590.57390.57860.5639
      MFPMC0.60420.61960.62560.62460.63750.6437
    • Table 5. Comparison of mAP values of ablation experiments

      View table

      Table 5. Comparison of mAP values of ablation experiments

      TaskMethodMIRFlickr-25KNUS-WIDE
      Image2TextBase0.72500.5630
      Base+IMFM0.73120.5734
      Base+TMFM0.73240.5727
      Base+IMFM+TMFM0.74010.5833
      Base+LFPMC0.75680.5974
      Base+IMFM+TMFM+LFPMC0.76870.6256
      Text2ImageBase0.73410.5605
      Base+IMFM0.73980.5769
      Base+TMFM0.74330.5834
      Base+IMFM+TMFM0.75220.6008
      Base+LFPMC0.76980.6294
      Base+IMFM+TMFM+LFPMC0.78980.6437
    Tools

    Get Citation

    Copy Citation Text

    Wanyu Deng, Yina Zhao, Wanzhen Yang, Bo Zhang, Hao Li, Shuqi Ye. Cross-Modal Hash Method Based on Multi-Scale Fusion and Projection Matching Constraint[J]. Laser & Optoelectronics Progress, 2022, 59(24): 2410006

    Download Citation

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

    Category: Image Processing

    Received: Sep. 17, 2021

    Accepted: Oct. 27, 2021

    Published Online: Oct. 31, 2022

    The Author Email: Zhao Yina (3065783275@qq.com)

    DOI:10.3788/LOP202259.2410006

    Topics