Opto-Electronic Engineering, Volume. 51, Issue 5, 240030(2024)

Real-time semantic segmentation algorithm based on BiLevelNet

Majing Wu1... Yong'ai Zhang1,2, Shanling Lin1,2, Zhixian Lin1,2, and Jianpu Lin12,* |Show fewer author(s)
Author Affiliations
  • 1School of Advanced Manufacturing, Fuzhou University, Quanzhou, Fujian 362200, China
  • 2Fujian Science & Technology Innovation Laboratory for Optoelectronic Information of China, Fuzhou, Fujian 350116, China
  • show less
    Figures & Tables(21)
    Network framework of BiLevelNet
    Comparison of different feature extraction modules
    AFR-S module and AFR module
    Schematic diagram of region perception and feature reuse
    BRA module
    Partial Bi-Level route attention module
    Decorder module
    Sample distribution of Cityscapes dataset and Camvid dataset
    Comparison of results with using the PBRA modules
    Comparison of segmentation results using bilinear interpolation
    Feature map before bilinear interpolation and the shallow feature map
    FADE upsampling segmentation results
    Visualization results of networks on Cityscapes dataset
    Visualization results of networks on the Camvid dataset
    • Table 1. Network framework of BiLevelNet

      View table
      View in Article

      Table 1. Network framework of BiLevelNet

      StageOperatorModeOutput size
      Stage 13 × 3 ConvStride 232 × 256 × 512
      3 × 3 ConvStride 132 × 256 × 512
      3 × 3 ConvStride 132 × 256 × 512
      Stage 2AFR-S64 × 128 × 256
      2 × ARFDilated 264 × 128 × 256
      Stage 3AFR-S128 × 64 × 128
      4 × AFRDilated 4128 × 64 × 128
      5 × AFRDilated 8128 × 64 × 128
      DecoderDAF32 × 256 × 512
      1 × 1 ConvStride 119 × 256 × 512
      Bilinear19 × 512 × 1024
    • Table 2. Performance comparison of different feature extraction modules on the Cityscapes dataset

      View table
      View in Article

      Table 2. Performance comparison of different feature extraction modules on the Cityscapes dataset

      Params/MFLOPs/GFPSmIoU/%
      SSnbt0.8311.6113267.1
      DAB0.7510.7814071.8
      AFR0.689.6412875.5
    • Table 3. Experimental results of different reduction factor modules in Cityscapes validation set

      View table
      View in Article

      Table 3. Experimental results of different reduction factor modules in Cityscapes validation set

      RatioParams/MFLOPs/GFPSmIoU/%
      00.679.5913574.0
      10.7410.2511674.2
      1/20.699.7512075.0
      1/40.689.6412875.5
      1/80.679.6113075.1
      1/160.679.613174.1
    • Table 4. Experimental results of FADE modules on the Cityscapes validation dataset

      View table
      View in Article

      Table 4. Experimental results of FADE modules on the Cityscapes validation dataset

      Params/MFLOPs/GFPSmIoU/%
      Bilinear0.689.6412875.5
      FADE0.710.412175.9
    • Table 5. Performance comparison of different models on the Cityscapes dataset

      View table
      View in Article

      Table 5. Performance comparison of different models on the Cityscapes dataset

      AlgorithmSizeParams/MFLOPs/GFPSmIoU/%
      ENet512×10240.364.354258.3
      ERFNet512×10242.1026.85968.0
      LEDNet512×10240.9411.57169.2
      DABNet512×10240.76-10470.1
      ELANet[29]512×10240.679.79374.7
      RELAXNet512×10241.9022.846474.8
      DALNet[30]512×1 0240.48-7471.1
      BiseNet-v2512×10243.4021.215672.6
      MIFNet[31]512×10240.8212.037473.1
      文献[32]512×10246.2212.5154.774.2
      Ours512×10240.7010.412175.1
    • Table 6. Evaluation results of per-class IoU /% on the Cityscapes dataset

      View table
      View in Article

      Table 6. Evaluation results of per-class IoU /% on the Cityscapes dataset

      ClassERFNetDABNetLEDNetFDDWNetOurs
      Roa97.996.897.198.098.0
      Sid82.178.578.682.482.2
      Bui90.790.990.491.191.8
      Wal45.245.346.552.554.8
      Fen50.450.148.151.256.5
      Pol59.059.160.959.963.2
      Tli62.665.260.464.468.4
      TSi68.470.771.168.972.1
      Veg91.992.591.292.592.8
      Ter69.468.160.070.370.5
      Sky94.294.693.294.494.5
      Ped78.580.574.380.882.3
      Rid59.858.551.859.865.2
      Car93.492.792.394.094.3
      Tru52.552.761.056.559.2
      Bus60.867.272.468.978.5
      Tra53.750.951.048.673.9
      Mot49.950.443.355.757.9
      Bic64.265.770.267.770.2
    • Table 7. Performance comparison on the CamVid dataset

      View table
      View in Article

      Table 7. Performance comparison on the CamVid dataset

      AlgorithmSizePretrainParams/MmIoU/%
      ENet360×480N0.3651.3
      CGNet360×480N0.564.7
      DALNet360×480N0.4766.1
      LEDNet360×480N0.9466.6
      DABNet360×480N0.7666.4
      MIFNet360×480N0.8167.7
      ELANet360×480N0.6767.9
      BiseNet-v2360×480Y5.868.7
      Ours360×480N0.768.2
    Tools

    Get Citation

    Copy Citation Text

    Majing Wu, Yong'ai Zhang, Shanling Lin, Zhixian Lin, Jianpu Lin. Real-time semantic segmentation algorithm based on BiLevelNet[J]. Opto-Electronic Engineering, 2024, 51(5): 240030

    Download Citation

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

    Category: Article

    Received: Jan. 30, 2024

    Accepted: Mar. 13, 2024

    Published Online: Jul. 31, 2024

    The Author Email: Lin Jianpu (林坚普)

    DOI:10.12086/oee.2024.240030

    Topics