• Laser & Optoelectronics Progress
  • Vol. 59, Issue 18, 1815011 (2022)
Bo Yang1, Zhenming Xu1、*, and Jianxin Liu2
Author Affiliations
  • 1China-UK Low Carbon College, Shanghai Jiao Tong University, Shanghai 201306, China
  • 2Nanjing Institute of Electronic Technology, Nanjing 210039, Jiangsu , China
  • show less
    DOI: 10.3788/LOP202259.1815011 Cite this Article Set citation alerts
    Bo Yang, Zhenming Xu, Jianxin Liu. Sorting and Detection of Impurity Glass Based on YOLOv4[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815011 Copy Citation Text show less
    Impurity glass samples. (a) Crystal; (b) glue; (c) laminated; (d) bar
    Fig. 1. Impurity glass samples. (a) Crystal; (b) glue; (c) laminated; (d) bar
    Three labelling methods for bar and copying. (a) Horizontal box; (b) rotated box; (c) square boxes; (d) small object copying
    Fig. 2. Three labelling methods for bar and copying. (a) Horizontal box; (b) rotated box; (c) square boxes; (d) small object copying
    Relationship between clustering effect of Kmeans++and k
    Fig. 3. Relationship between clustering effect of Kmeans++and k
    Clustering result of Kmeans++ algorithm
    Fig. 4. Clustering result of Kmeans++ algorithm
    YOLOv4 network structure and adjustment
    Fig. 5. YOLOv4 network structure and adjustment
    Loss and AP of validation set during training
    Fig. 6. Loss and AP of validation set during training
    Calculation method of occlusion proportion
    Fig. 7. Calculation method of occlusion proportion
    AP under different augmenting types before & after training
    Fig. 8. AP under different augmenting types before & after training
    Target pasting and detecting examples. (a)(b) Target attached with label; (c) interference; (d) occlusion
    Fig. 9. Target pasting and detecting examples. (a)(b) Target attached with label; (c) interference; (d) occlusion
    MethodAPmAP
    BarCrystalGlueLaminated
    YOLOv488.7797.3198.9698.0495.76
    YOLOv4+93.9698.2699.9798.0497.55
    YOLOV4_t+95.4798.9899.0797.9997.88
    Table 1. Accuracy comparison of different models in testset
    Type of boxesBarCrystalGlueLaminated
    GT262117110102
    TP25511610999
    FP1992221
    Table 2. Positive and negative sample distribution
    ParameterYOLOv4YOLOv3YOLOv4-tinyYOLOv3-tinyYOLOv4_t
    mAP0.960.920.880.860.98
    Speed/(frame·s-140.5438.51118.67158.2542.82
    Multi-Add/10929.7832.663.3928.96
    Params/106255.82246.1623.5217.90176.22
    Table 3. Parameter size and detection performance of typical networks
    Bo Yang, Zhenming Xu, Jianxin Liu. Sorting and Detection of Impurity Glass Based on YOLOv4[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815011
    Download Citation