• Acta Photonica Sinica
  • Vol. 51, Issue 2, 0210002 (2022)
Penghui ZHANG1, Zhi LIU2, Jianyong ZHENG3, Boxia HE1、*, and Yuhao PEI1
Author Affiliations
  • 1School of Mechanical Engineering,Nanjing University of Science and Technology,Nanjing 210094,China
  • 2Nanjing Planetech Intelligent Technology Co.,LTD,Nanjing 210014,China
  • 3Institute of Artificial Intelligence,Shanghai University,Shanghai 200444,China
  • show less
    DOI: 10.3788/gzxb20225102.0210002 Cite this Article
    Penghui ZHANG, Zhi LIU, Jianyong ZHENG, Boxia HE, Yuhao PEI. Real-time Infrared Target Detection Algorithm for Embedded System in Complex Scene[J]. Acta Photonica Sinica, 2022, 51(2): 0210002 Copy Citation Text show less
    The network structure of YOLOv4-Tiny
    Fig. 1. The network structure of YOLOv4-Tiny
    The network structure of YOLOv4-Tiny+SE+SPP/YOLOv4-Tiny+CBAM+SPP
    Fig. 2. The network structure of YOLOv4-Tiny+SE+SPP/YOLOv4-Tiny+CBAM+SPP
    Grad-CAM network visualization results
    Fig. 3. Grad-CAM network visualization results
    Quantity distribution of different categories in training and validation sets(left)test set(right)
    Fig. 4. Quantity distribution of different categories in training and validation sets(left)test set(right)
    Infrared and visible images of four different scenes
    Fig. 5. Infrared and visible images of four different scenes
    P-R curves of different categories on different networks
    Fig. 6. P-R curves of different categories on different networks
    Detection results of different models
    Fig. 7. Detection results of different models
    Grad-CAM in different networks visualization results
    Fig. 8. Grad-CAM in different networks visualization results
    ModelAP/%Precision/%Recall/%mAP/%FPS
    BicycleCarPerson
    YOLOv4-Tiny(original)32.8578.6561.8482.6941.3357.78232
    YOLOv4-Tiny+SE38.4780.7966.6180.3550.1861.95226
    YOLOv4-Tiny+CBAM38.9780.3466.7079.6550.4862.00215
    Table 1. Results of attentional mechanism ablation experiment
    ModelAP/%Precision/%Recall/%mAP/%FPS
    BicycleCarPerson
    YOLOv4-Tiny(original)32.8578.6561.8482.6941.3357.78232
    YOLOv4-Tiny+SE38.4780.7966.6180.3550.1861.95226
    YOLOv4-Tiny+SE+SPP42.7481.5668.6980.5453.0864.33212
    YOLOv4-Tiny+CBAM38.9780.3466.7079.6550.4862.00215
    YOLOv4-Tiny+CBAM+SPP42.0481.8969.5680.3353.0464.50202
    Table 2. Results of space pyramid pooling ablation experiment
    ModelAP/%Precision/%Recall/%mAP/%FPS
    BicycleCarPerson
    YOLOv4-Tiny(original)32.8578.6561.8482.6941.3357.78232
    ① YOLOv4-Tiny+SE+SPP42.7481.5668.6980.5453.0864.33212
    ① +Pretraining(VOC07+12)56.4584.3874.3882.2661.4771.74212
    ② YOLOv4-Tiny+CBAM+SPP42.0481.8969.5680.3353.0464.50202
    ② +Pretraining(VOC07+12)61.1284.4175.0583.5763.7473.53202
    Table 3. Results of ablation experiments on pre-trained models
    ModelDeviceAP/%Precision/%Recall/%mAP/%FPS
    BicycleCarPerson
    YOLOv4-Tiny(original)2 080 Ti32.8578.6561.8482.6941.3357.78232
    YOLOv4-Tiny+SE+SPP(our)56.4584.3874.3882.2661.4771.74212
    YOLOv4-Tiny+CBAM+SPP(our)61.1284.4175.0583.5763.7473.53202
    YOLOv4-Tiny(original)310 AI33.3278.4561.0584.5640.5757.6082
    YOLOv4-Tiny+SE+SPP(our)54.6183.0272.2783.1559.1769.9678
    YOLOv4-Tiny+CBAM+SPP(our)61.8784.1574.6184.8863.4673.5471
    Table 4. Experimental comparison results of embedded model deployment
    Penghui ZHANG, Zhi LIU, Jianyong ZHENG, Boxia HE, Yuhao PEI. Real-time Infrared Target Detection Algorithm for Embedded System in Complex Scene[J]. Acta Photonica Sinica, 2022, 51(2): 0210002
    Download Citation