Author Affiliations
1School of Electronic and Control Engineering, Chang'an University, Xi'an, Shaanxi 710064, China2Technology Quality Department, Shaanxi Province Railway Group Co., Ltd., Xi'an, Shaanxi 710199, China3Commission for Discipline Inspection and Supervision, Xi'an Xilan Natural Gas Group Co., Ltd., Xi'an, Shaanxi 710075, Chinashow less
Fig. 1. Different network models. (a) Classification model; (b) segmentation model
Fig. 2. Overall structure of the ResNet-GCN model. (a) Structure of the entire framework; (b) GCN structure; (c) boundary refinement module
Fig. 3. Dataset containing different crack types. (a) Crack; (b) watery crack; (c) crack with repair seal; (d) crack with lane line; (e) stitching seam; (f) crack containing debris
Fig. 4. Labeling of experimental crack data. (a) (b) (c) Original crack images; (d) (e) (f) manually marked cracks
Fig. 5. Comparision of GCN and ordinary convolution kernel
Fig. 6. Test accuracy of MobileNetv2-GCN model
Fig. 7. Test mIoU of MobileNetv2-GCN model
Fig. 8. Crack segmentation effect of MobileNetv2-GCN model. (a) Original images; (b) label images; (c) prediction results
Fig. 9. Crack skeleton extraction. (a)(b)(c) Binary images after segmentation; (d)(e)(f) extracted crack skeleton images
Fig. 10. Comparison of real and predicted average crack width pixel
Data | Training | Verification | Test |
---|
Image size /(pixel×pixel) | 512× 512 | 512×512 | 512×512 | Number of images | 2160 | 720 | 720 |
|
Table 1. Crack dataset details
k | Base | 3 | 5 | 7 | 9 | 11 | 13 | 15 |
---|
mIoU /% | 70.5 | 71.1 | 71.9 | 72.6 | 73.3 | 74.1 | 75.2 | 76.6 |
|
Table 2. Comparison of GCN results with different convolution kernel sizes
k | 3 | 5 | 7 | 9 | 11 |
---|
mIoU /% | GCN | 71.1 | 71.9 | 72.6 | 73.3 | 74.1 | Stack | 69.8 | 70.9 | 69.5 | 68.2 | 67.5 |
|
Table 3. Comparison of GCN and equivalent small kernel stack convolution
Number oflayers | Stack | GCN |
---|
2048 | 1024 | 210 | 2048 |
---|
mIoU/% | 71.1 | 70.5 | 68.6 | 72.8 | Parameteramount /k | 75885 | 28505 | 4307 | 608 |
|
Table 4. Comparison of experimental results to reduce the number of stacked convolution layers
Model | Boundaryaccuracy /% | Centeraccuracy /% | Overallaccuracy /% |
---|
Baseline | 71.3 | 90.1 | 70.3 | GCN | 71.5 | 91.1 | 85.6 | GCN+BR | 72.6 | 92.3 | 86.7 |
|
Table 5. Comparison of experimental results after adding boundary refinement blocks
Model | mIoU /% | Accuracy /% | Model size /MB |
---|
ResNet- GCN | 76.6 | 86.7 | 671.0 | ResNet50-GCN | 82.6 | 90.2 | 274.0 | ResNet101-GCN | 83.6 | 93.4 | 492.0 | MobileNetv2-GCN | 84.6 | 98.5 | 15.5 |
|
Table 6. Comparison of experimental results of different pre-training fusion models
Dataset | Crack500 | GAPs384 | Cracktree200 | Proposed |
---|
mIoU /% | 85.8 | 82.3 | 80.5 | 85.3 | Accuracy /% | 95.5 | 92.4 | 89.6 | 98.5 |
|
Table 7. Experimental results of MobileNetv2-GCN on different datasets
Model | SegNet | DeepLab | CNN | FCN | MobileNetv2-GCN |
---|
mIoU /% | 67.8 | 69.2 | 72.3 | 80.5 | 85.3 | Accuracy /% | 74.5 | 80.3 | 81.9 | 89.6 | 98.5 |
|
Table 8. Comparison of MobileNetv2-GCN and other crack segmentation models