• Laser & Optoelectronics Progress
  • Vol. 57, Issue 10, 101501 (2020)
Xiaohai Shen, Zehao Li, Min Li, Xiaolong Xu, and Xuewu Zhang*
Author Affiliations
  • College of Internet of Things Engineering, Hohai University, Changzhou, Jiangsu 213022, China
  • show less
    DOI: 10.3788/LOP57.101501 Cite this Article Set citation alerts
    Xiaohai Shen, Zehao Li, Min Li, Xiaolong Xu, Xuewu Zhang. Aluminum Surface-Defect Detection Based on Multi-Task Deep Learning[J]. Laser & Optoelectronics Progress, 2020, 57(10): 101501 Copy Citation Text show less
    Parameters sharing in multi-task deep learning
    Fig. 1. Parameters sharing in multi-task deep learning
    Multi-task deep neural network model of aluminum defect detection
    Fig. 2. Multi-task deep neural network model of aluminum defect detection
    Adaptive multi-task loss layer
    Fig. 3. Adaptive multi-task loss layer
    Aluminum defect image and ground truth. (a) Original image; (b) segmentation marked; (c) multi-label marked; (d) defect object marked
    Fig. 4. Aluminum defect image and ground truth. (a) Original image; (b) segmentation marked; (c) multi-label marked; (d) defect object marked
    Results of model output on test sets. (a) Original images; (b) segmentation; (c) defect detection; (d) defect multi-label detection
    Fig. 5. Results of model output on test sets. (a) Original images; (b) segmentation; (c) defect detection; (d) defect multi-label detection
    TaskTask weightMIoU(MPA) /%hloss /%mAP /%
    Seg.Clas.Obj.
    Segmentation10098.59(99.57)--
    Classification010-5.24-
    Object detection001--71.46
    Seg. +Clas.98.43(99.41)4.13-
    Seg. + Obj.98.82(99.73)-74.72
    Clas. + Obj.-1.0175.97
    Seg. +Clas. + Obj.98.57(99.48)0.9675.65
    Table 1. Performance of single task training and multi-task training on test sets
    MethodMIoU (MPA) /%hloss /%mAP /%
    FCN-8s98.75(99.49)--
    U-Net97.24(98.80)--
    ResNet50-6.41-
    ResNet101-4.67-
    Faster RCNN--69.78
    RetinaNet--72.13
    Proposed98.57(99.48)0.9675.65
    Table 2. Performance of several deep learning methods on test sets
    TaskAverageinference timeper image /msSum of singletask inferencetime /ms
    Segmentation14.960-
    Classification13.791-
    Object detection57.347-
    Seg. +Clas.15.43728.751
    Seg. + Obj.68.21272.307
    Clas. + Obj.66.02571.138
    Seg. +Clas. + Obj.69.43586.098
    Table 3. Inference time statics on single task model and multi-task model
    Xiaohai Shen, Zehao Li, Min Li, Xiaolong Xu, Xuewu Zhang. Aluminum Surface-Defect Detection Based on Multi-Task Deep Learning[J]. Laser & Optoelectronics Progress, 2020, 57(10): 101501
    Download Citation