• Laser & Optoelectronics Progress
  • Vol. 60, Issue 16, 1615007 (2023)
Xin Li, Xiangrong Li*, Cheng Wang, Qiuliang Li, and Zhuoyue Li
Author Affiliations
  • Fundamentals Department, Air Force Engineering University, Xi'an 710038, Shaanxi, China
  • show less
    DOI: 10.3788/LOP222557 Cite this Article Set citation alerts
    Xin Li, Xiangrong Li, Cheng Wang, Qiuliang Li, Zhuoyue Li. Aero-Engine Surface Defect Detection Model Based on Improved YOLOv5[J]. Laser & Optoelectronics Progress, 2023, 60(16): 1615007 Copy Citation Text show less
    YOLOv5s-6.0 network structure
    Fig. 1. YOLOv5s-6.0 network structure
    Flowchart of a search algorithm for data augmentation
    Fig. 2. Flowchart of a search algorithm for data augmentation
    Structure of the CA module
    Fig. 3. Structure of the CA module
    Diagram of CIoU
    Fig. 4. Diagram of CIoU
    Diagram of the CA module adding location
    Fig. 5. Diagram of the CA module adding location
    Type of defect. (a) Crack; (b) gap; (c) pit; (d) scratch
    Fig. 6. Type of defect. (a) Crack; (b) gap; (c) pit; (d) scratch
    Defect labelling example
    Fig. 7. Defect labelling example
    xml label file
    Fig. 8. xml label file
    Comparison of the detection effect of two models. (a) (b) (c) Detection effect of YOLOv5s; (d) (e) (f) Detection effect of YOLOv5-CE
    Fig. 9. Comparison of the detection effect of two models. (a) (b) (c) Detection effect of YOLOv5s; (d) (e) (f) Detection effect of YOLOv5-CE
    Data augmentation substrategyDescriptionRange of Vmagnitude
    ContrastAdjust the contrast of the image. Vmagnitude=0 gives a gray image and Vmagnitude=1 gives the original image[0,2]
    SharpnessAdjust the sharpness of the image. Vmagnitude=0 gives a blurred image and Vmagnitude=1 gives the original image[0,2]
    BrightnessAdjust the brightness of the image. Vmagnitude=0 gives a black image and Vmagnitude=1 gives the original image[0,2]
    RotationRotate the image by Vmagnitude degrees[-90°,90°]
    ScaleEnlarge or reduce the image to Vmagnitude scales[0.5,2]
    FlipFlip the imageFlip up-down/flip left-right
    HSV augmentationAdjust the H(hue),S(saturation),and V(value)of the imageH:[0°,360°],S:[0,1],V:[0,1]
    NoiseAdd noise to the imageGaussian noise and salt and pepper noise
    Table 1. Search space
    ModelPAP /%PmAP /%Speed /(frame/s)Capacity /MB
    crackgappitscratch

    Faster

    R-CNN

    75.478.158.783.173.814.29109
    YOLOv382.189.885.176.183.322.63235
    YOLOv490.093.286.679.087.224.72244
    YOLOv5s94.999.598.496.597.341.3214.4
    YOLOXs90.890.790.990.890.8110.6268.7
    YOLOv5-CE97.399.499.298.298.540.6514.5
    Table 2. Comparison of the detection performance of algorithms
    ModelRprecision /%Rrecall /%PAP /%PmAP /%Speed /(frame/s)
    crackgappitscratch
    YOLOv5s97.794.794.999.598.496.597.341.32
    YOLOv5_A96.895.795.299.598.997.897.940.49
    YOLOv5-C97.895.796.999.599.396.998.241.15
    YOLOv5-E97.396.696.499.599.398.298.446.73
    YOLOv5-CE98.195.497.399.499.298.298.540.65
    Table 3. Ablation experiments
    Xin Li, Xiangrong Li, Cheng Wang, Qiuliang Li, Zhuoyue Li. Aero-Engine Surface Defect Detection Model Based on Improved YOLOv5[J]. Laser & Optoelectronics Progress, 2023, 60(16): 1615007
    Download Citation