Optics and Precision Engineering, Volume. 32, Issue 22, 3395(2024)

Crowd counting method based on dense connection attention and scale perception recombination enhancement

Yong CHEN1,2、*, Ke DONG1, Zhuoaobo AN1, and Jianyu ZHOU1
Author Affiliations
  • 1School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou730070, China
  • 2Gansu Provincial Engineering Research Center for Artificial Intelligence and Graphics&Image Processing, Lanzhou730070, China
  • show less
    Figures & Tables(18)
    Overall network architecture diagram
    Densely connected convolutional attention network
    Attention comparison heat-map
    Multi-scale perception reorganization enhancement module
    Scale-aware reorganizing the upsampling architecture
    Soft mask selection mechanism
    Multi-resolution fusion structure
    Experimental results of ShanghaiTech dataset
    Experimental results of UCF-QNRF dataset
    Occlusion processing experiment results
    Experimental results of JHU_CROWD++ dataset
    • Table 1. Experimental results of different data centralization with heavy λ values

      View table
      View in Article

      Table 1. Experimental results of different data centralization with heavy λ values

      权重λSHASHBUCF-QNRFJHU_CROWD++
      MAEMSEMAEMSEMAEMSEMAEMSE
      1.065.1122.114.720.086.3177.774.1273.6
      0.963.2111.413.918.984.2156.372.9269.0
      0.862.7108.812.517.482.6155.470.2264.9
      0.761.0107.211.016.780.1154.068.5258.1
      0.659.4105.310.215.179.8153.866.9254.7
      0.558.6104.79.613.478.0152.364.2253.2
      0.457.2103.28.912.177.4148.162.1250.6
      0.356.8101.47.210.576.2146.759.6248.1
      0.256.1100.26.39.975.6145.058.3246.9
      0.155.499.65.89.374.3143.157.3245.3
    • Table 2. Experimental results of different models on ShanghaiTech

      View table
      View in Article

      Table 2. Experimental results of different models on ShanghaiTech

      方法SHASHB
      MAEMSEMAEMSE
      MCNN8110.5173.026.841.3
      CSRNet968.5115.410.916.2
      ACSCP1275.2102.117.027.1
      DM-Count2258.895.47.111.5
      DCANet2361.1108.28.415.0
      FIDTM2457.4103.97.312.0
      MP-Count2557.0103.17.212.4
      FFDB2656.6100.76.810.6
      本文方法55.499.65.89.3
    • Table 3. Experimental results of different models on UCF-QNRF

      View table
      View in Article

      Table 3. Experimental results of different models on UCF-QNRF

      方法UCF-QNRF
      MAEMSE
      Switch-CNN27228.4426.3
      TEDNet28112.9187.7
      DUBNet29105.8180.9
      DM-Count2285.2148.0
      DCANet2399.0177.3
      FIDTM2488.7153.1
      MP-Count2588.5151.3
      FFDB2678.1148.6
      本文方法74.3143.1
    • Table 4. Experimental results of different models on JHU_CROWD++

      View table
      View in Article

      Table 4. Experimental results of different models on JHU_CROWD++

      方法JHU_CROWD++
      MAEMSE
      MCNN8188.2482.7
      CRSNet986.1309.5
      CAN1199.4313.2
      DM-Count2268.2287.0
      DCANet2371.1177.4
      FIDTM2465.8253.1
      MP-Count2570.9260.7
      FFDB2659.2255.0
      本文方法57.3245.3
    • Table 5. Model quantitative index comparison

      View table
      View in Article

      Table 5. Model quantitative index comparison

      方法JHU_CROWD++
      Param/MGFLOPs/G
      MCNN80.1356.21
      CRSNet916.26857.84
      CAN1118.193.58
      DM-Count2221.560.8
      DCANet2335.2117.6
      FIDTM2456.770.34
      MP-Count2551.478.2
      FFDB2647.179.8
      本文方法23.342.49
    • Table 6. Comparison of model generalization

      View table
      View in Article

      Table 6. Comparison of model generalization

      模型部署方法MAEMSE
      JHU_CROWD++到SHAFIDTM2488.6150.0
      MP-Count2586.9145.5
      FFDB2684.2142.8
      本文方法66.1127.4
      JHU_CROWD++到SHBFIDTM2497.2195.5
      MP-Count2596.4192.0
      FFDB2695.9188.1
      本文方法84.7155.2
      JHU_CROWD++到UCF-QNRFFIDTM24106.2296.4
      MP-Count2594.7294.8
      FFDB2689.4291.0
      本文方法55.1223.8
    • Table 7. Comparison of ablation experimental indicators

      View table
      View in Article

      Table 7. Comparison of ablation experimental indicators

      Backbone模块MAEMSEParamGFLOs
      B1B2B3B4B5
      86.6314.18.922.87
      82.3301.210.225.42
      79.1298.311.629.7
      68.4277.515.833.1
      61.2255.418.339.5
      57.3245.323.342.49
    Tools

    Get Citation

    Copy Citation Text

    Yong CHEN, Ke DONG, Zhuoaobo AN, Jianyu ZHOU. Crowd counting method based on dense connection attention and scale perception recombination enhancement[J]. Optics and Precision Engineering, 2024, 32(22): 3395

    Download Citation

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

    Category:

    Received: Apr. 20, 2024

    Accepted: --

    Published Online: Mar. 10, 2025

    The Author Email: Yong CHEN (edukeylab@126.com)

    DOI:10.37188/OPE.20243222.3395

    Topics