Journals Articles Conferences News About CLP
Submit a paper Email Alert
Search
Search
Articles
Journals
News
Advanced Search
Top Searches
2D MaterialsTransformation opticsQuantum PhotonicsSmall lasersSemiconductor UV PhotonicsSupersymmetric microring laser arrays

Laser & Optoelectronics Progress, Volume. 62, Issue 14, 1415005(2025)

Lightweight Detection Network YOLOv10-EN Based on Public Safety Scenarios

Qixiang Meng1, Jingtao Wang1, Zhilin Gao1, Qiqi Kou2, and Fanliang Bu1、*
Author Affiliations
  • 1School of Information Network Security, People's Public Security University of China, Beijing 100038, China
  • 2School of Computer Science and Technology, China University of Mining and Technology, Xuzhou 221116, Jiangsu , China
    • show less
    AbstractGet PDF(in Chinese)
    Figures&Tables (23)
    Equations (12)
    References (35)
    Tools
    Paper Information
    Topics
    Figures & Tables(23)
    YOLOv10 network architecture

    Fig. 1. YOLOv10 network architecture

    Download full size
    YOLOv10-EN network

    Fig. 2. YOLOv10-EN network

    Download full size
    Calculation variables in regression loss function

    Fig. 3. Calculation variables in regression loss function

    Download full size
    Multi-scale convolution (MSConv) module

    Fig. 4. Multi-scale convolution (MSConv) module

    Download full size
    Structure of coordinate attention mechanism network

    Fig. 5. Structure of coordinate attention mechanism network

    Download full size
    Structure of EnC2f module

    Fig. 6. Structure of EnC2f module

    Download full size
    Samples in the dataset

    Fig. 7. Samples in the dataset

    Download full size
    Performance curves of YOLOv10n model based on different loss functions during the training phase. (a) Loss value; (b) F1 score; (c) mAP@0.5

    Fig. 8. Performance curves of YOLOv10n model based on different loss functions during the training phase. (a) Loss value; (b) F1 score; (c) mAP@0.5

    Download full size
    Ablation experimental results of MSConv module. (a) Detection legends of baseline model; (b) detection legends of MSConv optimization model; (c) heatmaps of baseline model; (d) heatmaps of MSConv optimization model

    Fig. 9. Ablation experimental results of MSConv module. (a) Detection legends of baseline model; (b) detection legends of MSConv optimization model; (c) heatmaps of baseline model; (d) heatmaps of MSConv optimization model

    Download full size
    Ablation experimental results of PSA module. (a) Detection legends of baseline model; (b) detection legends of PSA-CA optimization model; (c) heatmaps of baseline model; (d) heatmaps of PSA-CA optimization model

    Fig. 10. Ablation experimental results of PSA module. (a) Detection legends of baseline model; (b) detection legends of PSA-CA optimization model; (c) heatmaps of baseline model; (d) heatmaps of PSA-CA optimization model

    Download full size
    Composition of Bottleneck module

    Fig. 11. Composition of Bottleneck module

    Download full size
    Ablation experimental results of Bottleneck module. (a) Detection legends of baseline model; (b) detection legends of Bottleneck optimization model; (c) heatmaps of baseline model; (d) heatmaps of Bottleneck optimization model

    Fig. 12. Ablation experimental results of Bottleneck module. (a) Detection legends of baseline model; (b) detection legends of Bottleneck optimization model; (c) heatmaps of baseline model; (d) heatmaps of Bottleneck optimization model

    Download full size
    F1 score curves of each object detection model

    Fig. 13. F1 score curves of each object detection model

    Download full size
    P-R curves of each object detection model

    Fig. 14. P-R curves of each object detection model

    Download full size
    Experimental results of object detection models

    Fig. 15. Experimental results of object detection models

    Download full size

    • Table 1. Experimental conditions

      View table

      Table 1. Experimental conditions

      DeviceVersion
      GPURTX 3090
      CPUAMD Ryzen 7 5700X
      CUDA12.1
      CuDNN8.9.0.7

    • Table 2. Experimental parameter setting

      View table

      Table 2. Experimental parameter setting

      ParameterValue
      Image size640 pixel×640 pixel
      Learning rate0.01
      OptimizerAdam
      Batch size8
      Epoch100

    • Table 3. Ablation experimental settings for MSConv module

      View table

      Table 3. Ablation experimental settings for MSConv module

      No.MSConvGroupConvolution kernel size
      Group 1Group 2Group 3Group 4
      No.1√√None1×13×35×5
      No.2√√1×11×13×33×3
      No.3√√1×11×15×55×5
      No.4√√3×33×35×55×5
      No.5×（baseline）3×3
      No.6×（baseline）5×5

    • Table 4. Ablation experimental results for MSConv module

      View table

      Table 4. Ablation experimental results for MSConv module

      No.Params /103FLOPs /106F1 scoremAP@0.5
      No.1▲2615.59▲7769.08▲0.801▲0.799
      No.2★2658.94★7933.940.795★0.796
      No.32671.878064.47★0.7990.782
      No.42678.558130.120.7890.784
      No.5■2710.16■8433.12■0.784■0.774
      No.62816.649541.670.7800.776

    • Table 5. Ablation experimental results of PSA module

      View table

      Table 5. Ablation experimental results of PSA module

      AttentionParams /103FLOPs /106RatioF1 scoremAP@0.5
      0.51.01.5
      MHSA■2710.16■8433.12√■0.784■0.774
      3189.118818.22√0.7910.785
      4834.4210132.16√0.7940.788
      CA▲2622.54▲8352.04√★0.7970.789
      2994.838656.57√0.799★0.799
      4064.949500.94√▲0.806▲0.801

    • Table 6. Ablation experimental results of Bottleneck module

      View table

      Table 6. Ablation experimental results of Bottleneck module

      Convolution

      kernel

      		Params /103FLOPs /106ResidualF1 scoremAP@0.5
      （1）（2）（3）（4）（5）
      3×3 standard convolution■2710.16■8433.12√√√■0.784■0.774
      2710.168433.38√√√√★0.788★0.780
      2710.168433.41√√√√0.7870.778
      2710.168435.68√√√√√▲0.792▲0.783
      5×5 standard convolution■3930.64■12516.37√√√■0.778■0.772
      3930.6412516.94√√√√★0.7830.775
      3930.6412516.97√√√√0.781★0.778
      3930.6412517.57√√√√√▲0.786▲0.780

      Multi-scale-split-channel convolution

      （none， 1×1， 3×3， 5×5）

      		■2222.86■6848.98√√√■0.804■0.799
      2222.866849.07√√√√0.8070.802
      2222.866849.09√√√√★0.809★0.804
      2222.866849.18√√√√√▲0.813▲0.809

    • Table 7. Ablation experimental results of each optimization module

      View table

      Table 7. Ablation experimental results of each optimization module

      No.MSConvPSA-CABottleneckParams /103FLOPs /106F1 scoremAP@0.5
      MSConvResidual
      No.1■2710.16■8435.68■0.784■0.774
      No.2√2615.597769.080.8010.799
      No.3√2622.548352.040.7970.789
      No.4√√2222.866849.180.8130.809
      No.5√√2527.537690.160.8190.817
      No.6√√2188.686672.350.8060.803
      No.7√√2615.597771.830.8030.801
      No.8√√√2188.716674.110.8080.807
      No.9√√2196.556802.650.8150.812
      No.10√√2623.458353.180.8010.793
      No.11√√√2197.886804.130.8180.815
      No.12√√√▲2139.59▲6466.54★0.825★0.822
      No.13√√√2528.197692.110.8220.820
      No.14√√√√★2140.81★6468.42▲0.833▲0.824

    • Table 8. Performance comparison of object detection models

      View table

      Table 8. Performance comparison of object detection models

      ModelParams /106FLOPs /109

      F1

      score

      		mAP@0.5Latency /ms
      YOLOv5n［19］3.287.10.7660.7622.65
      YOLOv7［32］36.51103.30.7090.71414.4
      YOLOv8n［18］3.008.10.7710.7692.58
      YOLO-MS［31］4.138.240.7790.7723.46
      YOLOv9c［33］30.7101.70.7360.7258.13
      ■YOLOv10n［17］■2.69■8.4■0.784■0.774■2.42
      ★YOLOv11n［34］★2.65★6.5★0.788★0.778★2.39
      Faster R-CNN［1］28.33249.210.6640.625100.40
      SSD［3］24.4130.70.6670.62813.79
      RT-DETR［35］42.5136.50.7860.77712.66
      ▲Ours▲2.14▲6.47▲0.833▲0.824▲2.35
    Tools

    Get Citation

    Copy Citation Text

    Qixiang Meng, Jingtao Wang, Zhilin Gao, Qiqi Kou, Fanliang Bu. Lightweight Detection Network YOLOv10-EN Based on Public Safety Scenarios[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1415005

    Download Citation

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

    Category: Machine Vision

    Received: Jan. 3, 2025

    Accepted: Feb. 5, 2025

    Published Online: Jul. 3, 2025

    The Author Email: Fanliang Bu (20051257@ppsuc.edu.cn)

    DOI:10.3788/LOP250445

    CSTR:32186.14.LOP250445

    Topics

    laser devices and laser physics
    Lasers and Laser Optics
    Laser physics
    laser manufacturing
    Instrumentation, Measurement and Metrology