• Journal of Applied Optics
  • Vol. 45, Issue 4, 732 (2024)
Kai WANG, Shuli LOU*, and Yan WANG
Author Affiliations
  • School of Physics and Electronic Information, Yantai University, Yantai 264005, China
  • show less
    DOI: 10.5768/JAO202445.0402002 Cite this Article
    Kai WANG, Shuli LOU, Yan WANG. Small object detection algorithm based on improved YOLOv3[J]. Journal of Applied Optics, 2024, 45(4): 732 Copy Citation Text show less
    Structure diagram of YOLOv3 network
    Fig. 1. Structure diagram of YOLOv3 network
    Diagram of Mosaic, Mixup and combination data enhancement example
    Fig. 2. Diagram of Mosaic, Mixup and combination data enhancement example
    Structure diagram of improved YOLOv3 network
    Fig. 3. Structure diagram of improved YOLOv3 network
    Schematic diagram of CSP module, SimSPPF module and feature enhancement module
    Fig. 4. Schematic diagram of CSP module, SimSPPF module and feature enhancement module
    Three cases with the same IoU value
    Fig. 5. Three cases with the same IoU value
    Comparison of detection effect of two algorithms on multi-target images
    Fig. 6. Comparison of detection effect of two algorithms on multi-target images
    配置深度学习框架操作系统CPUGPU运行内存CUDA
    型号Pytorch 1.13Windows 10Inter Core i7-7770Tesla P10016 GB11.3
    Table 1. Experimental environment configuration
    算法类别主干网络mAP/%FPSGPU
    SSD 300VGG1678.5371Tesla P100
    YOLOv3Darknet5381.3449Tesla P100
    文献[14]算法Darknet5383.2622RTX 2070
    文献[15]算法Darknet5383.832TITAN RTX
    文献[16]算法VGG1679.940GTX 1080Ti
    文献[17]算法Darknet5384.619RTX 2080Ti
    YOLOv4CSPDarknet5385.251Tesla P100
    本文算法Darknet5385.7446Tesla P100
    Table 2. Comparison of detection effects by different algorithms
    类别SSDYOLOv3本文算法
    sofa22.1823.818.4
    bicycle3131.717.2
    motorbike21.524.316
    dog2922.714.7
    sheep3339.427.3
    horse16.117.511.5
    Table 3. Comparison of missed detection rates for object detection %
    类型mAPR
    小 (0,32]中 (32,96]大 (96,416]小 (0,32]中 (32,96]大 (96,416]
    SSD8.923.55424.638.665.1
    YOLOv310.725.345.326.142.156.3
    YOLOv417.54357.335.355.466.7
    本文算法19.43559.241.64868.1
    Table 4. Effect comparison of different algorithms on different scale objects %
    组别算法改进方式mAP/%FPS
    小尺度总体
    1YOLOv332.781.3449
    2YOLOv3+数据增强34.381.9149
    3YOLOv3+数据增强+GIoU37.183.0248
    4YOLOv3+数据增强+GIoU+改进特征融合网络42.584.6446
    5YOLOv3+数据增强+GIoU+改进特征融合网络+特征增强模块46.385.7446
    Table 5. Results of ablation comparison experiment