Author Affiliations
1School of Electronics and Control Engineering, Chang’an University, Xi’an 710064, Shaanxi , China2Information Department, Guangxi New Development Transportation Group Co., Ltd., Nanning 530029, Guangxi , Chinashow less
Fig. 1. Depth separable convolution. (a) Convolution process; (b) convolution structure
Fig. 2. Pyramid pooling module
Fig. 3. Improved multi-branch feature shared structure network
Fig. 4. Structure diagram of modules. (a) GCN module; (b) BR module
Fig. 5. Structure diagram of RRC module
Fig. 6. Process of dataset annotation and crop
Fig. 7. Training loss and accuracy obtained by changing the initial learning rate. (a) Loss; (b) accuracy
Fig. 8. Visual comparison results of different algorithms in crack images with noise
Fig. 9. Change trend of each indicator. (a) Evaluation index; (b) PR curve
Fig. 10. Segmentation effects of the proposed algorithm on the public dataset
Fig. 11. Schematic of visualized results of crack skeleton extraction
Fig. 12. Comparison of measurement errors of crack length and width
Parameter | Training set | Validation set | Test set |
---|
Image size /(pixel×pixel) | | | | Number of images /frame | 12000 | 4000 | 4000 |
|
Table 1. Partition of crack dataset
Algorithm | Precision | Recall | F1 | MIoU | FPS(Millisecond/image) |
---|
Fast-SCNN | 0.7548 | 0.7962 | 0.7749 | 0.7692 | 67.5(14.81) | SegNet | 0.8312 | 0.7864 | 0.8082 | 0.8206 | 18.7(53.48) | Pspnet | 0.9021 | 0.9115 | 0.9068 | 0.8021 | 29.3(34.13) | DeepCrack | 0.8463 | 0.8159 | 0.8308 | 0.8642 | 25.6(39.06) | CrackU-Net | 0.8736 | 0.7938 | 0.8318 | 0.8527 | 34.8(28.74) | Proposed algorithm | 0.9695 | 0.9283 | 0.9485 | 0.8902 | 59.4(16.84) |
|
Table 2. Performance comparison of different crack detection algorithms
Dataset | Precision | Recall | F1 | MIoU | FPS(Millisecond/image) |
---|
Cracktree200 | 0.9432 | 0.9509 | 0.9470 | 0.8772 | 54.5(18.35) | GAPs384 | 0.9527 | 0.9445 | 0.9486 | 0.8495 | 53.8(18.59) | Crack500 | 0.9683 | 0.9485 | 0.9583 | 0.8679 | 56.2(17.79) |
|
Table 3. Test results of the proposed algorithm on different public datasets
Parameter | Fast-SCNN | SegNet | Pspnet | DeepCrack | CrackU-Net | Proposed algorithm |
---|
Relative error of length /% | 8.62 | 7.28 | 6.96 | 5.73 | 5.14 | 4.73 | Relative error of width /% | 7.94 | 6.53 | 7.32 | 6.87 | 6.02 | 5.21 |
|
Table 4. Relative errors of length and width of different comparison algorithms