1 Introduction
2 Related Work
3 Multi-Scale Splicing Attention U-Net
3.1 Preliminaries
3.2 Construction of the Proposed Network
4 Methodology for Gear Pitting Measurement
4.1 Pitting Dataset Construction and Preprocessing
4.2 Tooth Surface and Pitting Segmentation
5 Experimental Results
5.1 Operating Environment and Evaluation Indicators
Hyper parameter | Value |
---|---|
Learning rate | 0.0025 |
Optimization algorithm | SGD |
Multiple of learning rate change | 0.1 |
Step of learning rate change | 20000 |
Rotation_range | 0.2 |
Height_shift_range | 0.05 |
Width_shift_range | 0.05 |
Zoom_range | 0.05 |
Shear_range | 0.05 |
Steps_per_epoch | 1000 |
Epochs | 400 |
Batch size | 2 |
5.2 Dataset Introduction
Name | Number of images | Number of testing images | Resolution | Minimum object size | Maximum object size |
---|---|---|---|---|---|
Gear pitting | 2200 | 200 | 1280 × 1024 | 15 × 15 pixels | 1213 × 235 pixels |
Crackforest | 1500 | 300 | 480 × 320 | 57 × 37 pixels | 480 × 246 pixels |
Pascal VOC | 13487 | 1456 | 500 × 375 | 54 × 43 pixels | 489 × 334 pixels |
5.3 Comprehensive Evaluation Indexes
5.4 Results Visualization
Model | Re | P | R | MIoU |
---|---|---|---|---|
U-Net | 7.80 | 86.93 | 90.56 | 79.80 |
Attention U-Net | 10.55 | 87.84 | 83.11 | 74.41 |
MSSA U-Net | 6.23 | 87.97 | 89.82 | 80.17 |
Model | Training time (h) | Running time (s) |
---|---|---|
U-Net | 12.10 | 2.95 |
Attention U-Net | 13.58 | 3.42 |
MSSA U-Net | 15.44 | 3.71 |
Model | P | R | MIoU | F1 |
---|---|---|---|---|
U-Net | 90.37 | 89.74 | 85.01 | 90.05 |
Attention U-Net | 88.16 | 86.10 | 82.43 | 87.12 |
MSSA U-Net | 93.61 | 91.02 | 88.75 | 92.29 |
Model | P | R | MIoU | F1 |
---|---|---|---|---|
U-Net | 80.81 | 79.77 | 70.85 | 78.38 |
Attention U-Net | 77.48 | 76.49 | 67.72 | 75.51 |
MSSA U-Net | 82.54 | 81.80 | 73.43 | 80.40 |