Author Affiliations
1Beijing Key Lab of Traffic Data Analysis and Mining, School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, China2School of Computer and Information Technology, Beijing Jiaotong University, Beijing 100044, Chinashow less
Fig. 1. Semi-supervised training framework
Fig. 2. MSCM structure
Fig. 3. Improved network framework
Fig. 4. Pseudo tags obtained by different methods. (a) Original images; (b) ground-truth; (c) SF method; (d) wCtr method; (e) GC method
Fig. 5. Crack-detection effect under different networks. (a) Original images; (b) ground-truth; (c) SegNet network; (d) DeepCrack network; (e) proposed network
Fig. 6. Cack detection effect after training with different proportion of manual tag and pseudo tag. (a) Original images; (b) ground-truth; (c) 0; (d) 1/150; (e) 1/65; (f) 1/30; (g) 1/15; (h) 1/6; (i) 1
Fig. 7. Detection effect of different cracks in different networks under full supervision. (a) Original images; (b) ground-truth; (c) SegNet network; (d) DeepCrack network; (e) proposed network
Network | Precision /% | Recall /% | F1-score |
---|
SegNet[19] | 85.23 | 89.77 | 86.74 | DeepCrack[14] | 81.71 | 92.88 | 86.26 | Proposed network | 80.29 | 93.56 | 85.66 |
|
Table 1. Crack-detection results of different networks
Method | Precision /% | Recall /% | F1-score |
---|
SF[16] | 76.87 | 72.37 | 69.22 | wCtr[17] | 74.33 | 71.30 | 67.34 | GC[18] | 86.57 | 66.93 | 70.00 |
|
Table 2. Quantitative analysis results of three methods
Proportion | Precision /% | Recall /% | F1-score |
---|
0 | 72.93 | 85.21 | 74.52 | 1/150 | 72.74 | 88.62 | 76.59 | 1/65 | 77.48 | 88.50 | 80.47 | 1/30 | 77.68 | 92.67 | 83.39 | 1/15 | 80.29 | 93.56 | 85.66 | 1/6 | 85.99 | 92.17 | 88.42 | 1 | 89.34 | 97.48 | 92.31 |
|
Table 3. Quantitative analysis results of manual labeled dataset and pseudo labeled dataset in different proportions
Proportion | Precision /% | Recall /% | F1-score |
---|
1/150 | 14.26 | 50.35 | 21.46 | 1/65 | 25.85 | 60.38 | 35.04 | 1/30 | 49.83 | 63.80 | 54.81 | 1/15 | 69.48 | 83.51 | 75.21 | 1/6 | 82.16 | 91.59 | 86.22 |
|
Table 4. Quantitative analysis results after pre-training on different proportions of manually annotated datasets
Fusion coefficient | Precision /% | Recall /% | F1-score |
---|
1∶0∶0 | 81.78 | 89.81 | 84.81 | 0∶1∶0 | 81.07 | 88.14 | 83.68 | 0∶0∶1 | 85.32 | 95.03 | 89.34 | 1∶1∶1 | 80.29 | 93.56 | 85.66 | 1∶1∶2 | 85.93 | 91.33 | 87.69 | Union | 80.39 | 91.67 | 84.77 |
|
Table 5. Crack detection results of SF, wCtr and GC methods under different fusion coefficients
Network | Precision /% | Recall /% | F1-score |
---|
SegNet [19] | 85.23 | 89.77 | 86.74 | DeepCrack[14] | 81.71 | 92.88 | 86.26 | Proposed network | 89.34 | 96.48 | 93.31 |
|
Table 6. Crack detection results of different networks under full supervision