• Laser & Optoelectronics Progress
  • Vol. 60, Issue 4, 0415005 (2023)
Yishan Dong1,1,">, Zhaoxin Li1,1,">, Jingyuan Guo1,1,">, Tianyu Chen1,1,">, and Shuhua Lu1,1,2,">*
Author Affiliations
  • 1College of Information and Cyber Security, People's Public Security University of China, Beijing 102600, China
  • 2Key Laboratory of Security Technology and Risk Assessment Ministry of Public Security, Beijing 102600, China
  • show less
    DOI: 10.3788/LOP212848 Cite this Article Set citation alerts
    Yishan Dong, Zhaoxin Li, Jingyuan Guo, Tianyu Chen, Shuhua Lu. Improved YOLOv5 Model for X-Ray Prohibited Item Detection[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0415005 Copy Citation Text show less
    Network structure. (a) Improved YOLOv5 model; (b) FPN and PAN
    Fig. 1. Network structure. (a) Improved YOLOv5 model; (b) FPN and PAN
    Structure of CBAM attention module
    Fig. 2. Structure of CBAM attention module
    Schematic diagrams of WBF and NMS algorithms
    Fig. 3. Schematic diagrams of WBF and NMS algorithms
    Example of Mixup data augmentation
    Fig. 4. Example of Mixup data augmentation
    Detection results of the two models before and after improvement on the SIXray dataset
    Fig. 5. Detection results of the two models before and after improvement on the SIXray dataset
    Confusion matrix and P-R curve of proposed method on SIXray, HiXray, OPIXray dataset. FP for background false positive. (a) Confusion matrix on SIXray; (b) confusion matrix on OPIXray; (c) confusion matrix on HiXray; (d) P-R curves on SIXray; (e) P-R curves on OPIXray; (f) P-R curves on HiXray
    Fig. 6. Confusion matrix and P-R curve of proposed method on SIXray, HiXray, OPIXray dataset. FP for background false positive. (a) Confusion matrix on SIXray; (b) confusion matrix on OPIXray; (c) confusion matrix on HiXray; (d) P-R curves on SIXray; (e) P-R curves on OPIXray; (f) P-R curves on HiXray
    Example of bounding boxes which are taken from SIXray dataset. (a) Original images ; (b) bounding boxes generated by NMS; (c) bounding boxes generated by WBF
    Fig. 7. Example of bounding boxes which are taken from SIXray dataset. (a) Original images ; (b) bounding boxes generated by NMS; (c) bounding boxes generated by WBF
    MethodmAP /%G /%K /%W /%P /%S /%Size /MB
    YOLOv5s87.298.179.284.890.383.814.04
    Ours(5s)89.698.482.288.492.286.715.72
    YOLOv5m89.598.283.187.694.284.541.46
    Ours(5m)92.599.185.690.595.292.045.18
    Table 1. Comparison indicators of the two models before and after improvement
    MethodmAP /%
    YOLOv32979.2
    YOLOv43083.1
    ASPP-YOLOv43185.2
    SSD3282.9
    YOLOv5s2687.2
    YOLOv5s+Ours89.6
    Table 2. Comparison results on the SIXray dataset
    Method

    OPIXray

    mAP /%

    HiXray

    mAP /%

    YOLOv32978.2-
    YOLOv3+DOAM2079.2-
    YOLOv43078.9-
    CHR1878.6-
    FBS1981.7-
    SSD3270.971.4
    SSD+DOAM2074.072.1
    SSD+LIM2874.673.1
    FCOS3382.075.7
    FCOS+DOAM2082.476.2
    FCOS+LIM2883.177.3
    YOLOv5s2687.881.7
    YOLOv5s+DOAM2088.082.2
    YOLOv5s+LIM2890.683.2
    YOLOv5s+Ours91.683.1
    Table 3. Comparison results on the OPIXray and HiXray datasets
    ComponentYOLOv5mYOLOv5s
    mAP /%Size /MBmAP /%SIZE /MB
    Baseline89.541.4687.214.04
    Baseline+CBAM89.745.1888.615.72
    Baseline+Mixup90.441.4687.614.04
    Baseline+Mixup+CBAM90.545.1888.015.72
    Baseline+WBF90.241.4688.214.04
    Baseline+WBF+CBAM90.145.1889.115.72
    Baseline+WBF+Mixup91.041.4689.414.04
    Baseline+WBF+Mixup +CBAM92.545.1889.615.72
    Table 4. Ablation study on the SIXray dataset
    Yishan Dong, Zhaoxin Li, Jingyuan Guo, Tianyu Chen, Shuhua Lu. Improved YOLOv5 Model for X-Ray Prohibited Item Detection[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0415005
    Download Citation