Author Affiliations
1State Key Laboratory of Precision Measuring Technology and Instrument, Tianjin University, Tianjin 300072, China2Labotatory of Micro/Nano Manufacturing Technology (MNMT), Tianjin 300072, Chinashow less
Fig. 1. Data collection system
Fig. 2. Overall defect detection scheme based on surface analysis
Fig. 3. Schematic diagram of intelligent profile analysis
Fig. 4. Structure of FETN
Fig. 5. Structure of FIFE block
Fig. 6. Structure of cascaded receptive field enhancement defect detection model
Fig. 7. Backbone network structure of defect detection model. (a) Backbone network structure; (b) comparison of defect characteristic receptive field
Fig. 8. UPP-CLS dataset label distribution. (a) Label distribution before data balancing; (b) label distribution after data balancing
Fig. 9. Results of FETN intelligent profile analysis and filtering
Fig. 10. Height-to-width ratio statistics of dimension box in dataset
Fig. 11. Comparison of the results between the proposed detection model and other mainstream detection models on the UPP-DET dataset. (a) Overall comparison; (b) partial comparison
Fig. 12. Test results of defect detection model
Parameter | UPP-CLS | UPP-DET |
---|
Batch size | 8 | 4 | Epoch | 50 | 15 | Initial learning rate | 0.001 | 0.0002 | Weight decay | ‒ | 0.05 | Optimizer | Adam | AdamW | Regularization | L2 weight decay | ‒ |
|
Table 1. Basic parameters of the experiment
Vector ordering | SC | SE | Accuracy /% |
---|
L | C |
---|
‒ | ‒ | ‒ | ‒ | 75.18 | √ | ‒ | ‒ | ‒ | 74.39 | ‒ | √ | ‒ | ‒ | 78.07 | ‒ | √ | √ | ‒ | 78.24 | ‒ | √ | √ | √ | 79.47 |
|
Table 2. Module performance analysis in the FIFE Block
Image branch backbone | FIFE | Accuracy /% | Sensitivity /% | FNR /% | TNR /% | Specificity /% |
---|
InceptionV3[23] | ‒ | 70.47 | 68.83 | 31.17 | 73.17 | 26.83 | √ | 74.44 | 71.23 | 28.77 | 77.70 | 22.30 | ResNet-50[21] | ‒ | 75.18 | 73.64 | 26.36 | 77.88 | 22.12 | √ | 79.47 | 76.12 | 23.88 | 76.27 | 23.73 | ResNet-101[21] | ‒ | 78.63 | 76.03 | 23.97 | 80.00 | 20.00 | √ | 83.98 | 79.68 | 20.32 | 83.21 | 16.79 | MobileNetV3[24] | ‒ | 70.09 | 72.11 | 27.89 | 71.67 | 28.33 | √ | 75.69 | 72.99 | 27.01 | 76.35 | 23.65 | ResNext-50[25] | ‒ | 78.87 | 77.69 | 22.31 | 80.04 | 19.96 | √ | 82.21 | 77.15 | 22.85 | 84.76 | 15.24 | EfficientNet-b2[19] | ‒ | 80.41 | 78.25 | 21.75 | 82.67 | 17.33 | √ | 84.86 | 80.89 | 19.11 | 86.38 | 13.62 | EfficientNet-b4[19] | ‒ | 82.11 | 79.97 | 20.03 | 84.64 | 15.36 | √ | 85.36 | 81.49 | 18.51 | 87.72 | 12.28 |
|
Table 3. Comparative analysis of each model embedded in FIFE block
AD | Mixup | DC | mAP |
---|
— | — | — | 0.697 | √ | — | — | 0.721 | √ | √ | — | 0.729 | √ | √ | √ | 0.740 |
|
Table 4. Performance analysis of each model embedded in FIFE block
Stage | Method | Backbone | MST | Scratch | Pit | mAP | FPS /(frame·s-1) |
---|
One-stage | SSD300[27] | VGG-16 | ‒ | 0.363 | 0.335 | 0.349 | ‒ | SSD512[27] | VGG-16 | ‒ | 0.581 | 0.593 | 0.587 | ‒ | RetinaNet[28] | ResNet-50 | ‒ | 0.775 | 0.749 | 0.762 | ‒ | Yolov3[29] | DarkNet-53 | ‒ | 0.632 | 0.594 | 0.613 | ‒ | Yolov5 | DarkNet-53 | ‒ | 0.816 | 0.782 | 0.799 | ‒ | Two-stage | Faster R-CNN[30] | ResNet-50 | √ | 0.729 | 0.751 | 0.740 | 28.3 | Dynamic R-CNN[31] | ResNet-50 | √ | 0.816 | 0.798 | 0.807 | 24.9 | Cascade R-CNN[18] | ResNet-50 | √ | 0.815 | 0.822 | 0.819 | 22.7 | Proposed | EfficientNet-b4 | √ | 0.867 | 0.840 | 0.854 | 21.1 |
|
Table 5. Comparison results between proposed model and other detection models
No. | | mAP |
---|
Baseline | [0.5,0.6,0.7] | 0.854 | 1 | [0.2,0.6,0.7] | 0.841 | 2 | [0.3,0.6,0.7] | 0.853 | 3 | [0.4,0.6,0.7] | 0.856 | 4 | [0.4,0.5,0.6] | 0.851 | 5 | [0.4,0.5,0.7] | 0.862 | 6 | [0.4,0.5,0.8] | 0.839 |
|
Table 6. IoU threshold analysis results at each stage of defect detection model detection head