Acta Optica Sinica, Volume. 45, Issue 9, 0915001(2025)

Surface Defect Recognition of Laser Cladding Layer Based on Deep Learning

Da Chen, Xuan Zhang, Shengbin Zhao*, and Mingdi Wang**
Author Affiliations
  • School of Mechanical and Electrical Engineering, Soochow University, Suzhou 215131, Jiangsu , China
  • show less
    Figures & Tables(16)
    Structure of experimental platform
    Relative relationship of visual system
    Cladding system. (a) Preset powder; (b) powder spraying
    Three types of cladding state. (a) Good cladding; (b) uneven cladding; (c) not fused
    Model architectures. (a) Traditional ResNet architecture; (b) dual-channel ResNet architecture
    Image transformations used for augmentation. (a) Original image; (b) rotation; (c) random aspect ratio; (d) flipping
    Defect imaging enhancement process. (a) Original image; (b) removing noise; (c) dictionary matrix in K-SVD algorithm; (d) final image
    Flowchart of K-SVD
    Curves of accuracy and loss. (a) Accuracy curve; (b) loss curve
    Feature maps of ResNet features from different layers
    Rendering of Guided Grad-CAM and its flow chart for good cladding images
    • Table 1. Acquisition parameters of vision system

      View table

      Table 1. Acquisition parameters of vision system

      ParameterValue
      Speed55 frame/s
      Resolution1280 pixel×1024 pixel
      Dynamic range>140 dB
      Shutter range1 μs‒53 s exposure
    • Table 2. Three types of machining parameters

      View table

      Table 2. Three types of machining parameters

      Cladding stateLaser power /WPowder gas flow rate /(L/min)Cladding line speed /(mm/s)Cladding spacing /mm
      Good cladding25007.5‒8.015001.2
      Uneven cladding2000‒22007.5‒8.020001.2
      Not fused300013.522001.2
    • Table 3. Comparison of algorithms

      View table

      Table 3. Comparison of algorithms

      IndicatorWiener filterMedian filterK-SVD
      PSNR /dB28.526.832.1
      SSIM0.820.750.92
      Time taken /s3.05.55.2
      Noise typeGaussian noiseSalt-and-pepper noiseGaussian & salt-and-pepper noise
      AdaptabilityEdge modelImage modelEdge model
    • Table 4. Comparison of models

      View table

      Table 4. Comparison of models

      Cladding stateCenter loss-ResNetSVM
      PrecisionAccuracyPrecisionAccuracy
      Not fused1.0001.0001.0000.275
      Good cladding0.9850.9480.5450.614
      Uneven cladding0.9680.9860.9640.915
    • Table 5. Comparison of old and new algorithms

      View table

      Table 5. Comparison of old and new algorithms

      Cladding stateCenter loss-ResNetResNet
      PrecisionAccuracyPrecisionAccuracy
      Not fused1.0001.0000.9750.992
      Good cladding0.9850.9481.0000.967
      Uneven cladding0.9680.9860.9841.000
    Tools

    Get Citation

    Copy Citation Text

    Da Chen, Xuan Zhang, Shengbin Zhao, Mingdi Wang. Surface Defect Recognition of Laser Cladding Layer Based on Deep Learning[J]. Acta Optica Sinica, 2025, 45(9): 0915001

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Machine Vision

    Received: Jan. 8, 2025

    Accepted: Mar. 11, 2025

    Published Online: May. 16, 2025

    The Author Email: Shengbin Zhao (shengbinz@163.com), Mingdi Wang (wangmingdi@suda.edu.cn)

    DOI:10.3788/AOS250462

    CSTR:32393.14.AOS250462

    Topics