• Laser & Optoelectronics Progress
  • Vol. 58, Issue 14, 1410009 (2021)
Junfeng Li1、*, Yansen He1, and Wenzhan Dai2
Author Affiliations
  • 1School of Mechanical Engineering and Automation, Zhejiang Sci-Tech University, Hangzhou, Zhejiang 310018, China
  • 2School of Information and Electronic Engineering, Zhejiang Gongshang University, Hangzhou, Zhejiang 310018, China
  • show less
    DOI: 10.3788/LOP202158.1410009 Cite this Article Set citation alerts
    Junfeng Li, Yansen He, Wenzhan Dai. Light Guide Plate Defect Detection Combing Light Weight and Cascade Deep Learning Network[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410009 Copy Citation Text show less
    Partial images of light guide plate
    Fig. 1. Partial images of light guide plate
    Various defects on light guide plate. (a) Dot defect; (b) line defect; (c) area defect
    Fig. 2. Various defects on light guide plate. (a) Dot defect; (b) line defect; (c) area defect
    Structure of defect detection of light guide plate
    Fig. 3. Structure of defect detection of light guide plate
    Replacement and fusion of defect background
    Fig. 4. Replacement and fusion of defect background
    Structure of LW_CNN
    Fig. 5. Structure of LW_CNN
    Lightweight module
    Fig. 6. Lightweight module
    Structure of IR_CNN
    Fig. 7. Structure of IR_CNN
    Experimental device. (a) Three-dimensional device; (b) physical device
    Fig. 8. Experimental device. (a) Three-dimensional device; (b) physical device
    Training curve
    Fig. 9. Training curve
    Visualization of partial feature map. (a) Normal; (b) dot defect; (c) line defect; (d) area defect
    Fig. 10. Visualization of partial feature map. (a) Normal; (b) dot defect; (c) line defect; (d) area defect
    DatasetKindNumberTrainingValidationTest
    Dataset1Positive(normal)155001085015503100
    Negative(defective)11373796211382273
    Dataset2Normal40002800400800
    Area37062595371740
    Lines39672777397793
    Dots41972938420839
    Table 1. Dataset setting
    Network2-classes4-classesFLOPs /GSpeed /s-1
    Normal /%Defects /%Normal /%Area /%Lines /%Dots /%
    VGG19[21]97.5497.2197.2597.0397.2297.7419.61232
    ResNet18[23]99.5899.5299.5098.9298.8798.331.74427
    MobileNetv2[25]98.8699.3399.4697.3396.6898.490.06660
    LW_CNN99.6599.6398.5293.3298.2595.790.02811
    IR_CNN99.7799.7899.6399.3298.9999.177.63320
    Table 2. Test results of different classification networks on light guide plate dataset
    Defect typeNumberAccuracy of proposed method /%Accuracy of method in Ref. [18] /%Speed of proposed method /s-1Speed of method in Ref. [18] /s-1
    Normal150100.0096.001.956.8
    Area defect8598.8287.06
    Line defect12796.8596.85
    Dot defect13897.8395.65
    Average50098.4094.60
    Table 3. Comparision of algorithm in this paper and method in Ref. [18]
    Junfeng Li, Yansen He, Wenzhan Dai. Light Guide Plate Defect Detection Combing Light Weight and Cascade Deep Learning Network[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410009
    Download Citation