Fig. 1. Different types of bridge cracks. (a) Single small crack; (b) single thick crack; (c) cross crack; (d) reticulated crack
Fig. 2. Partial images of expanded bridge crack image dataset. (a) Original images; (b) 90° rotation; (c) 180° rotation; (d) horizontal flip; (e) vertical flip
Fig. 3. Structure of multi-resolution network model
Fig. 4. Schematic of fusion operation. (a) Low and medium resolution merged into high resolution; (b) high and low resolution merge into medium resolution; (c) medium and high resolution merge into low resolution
Fig. 5. HRBCS model
Fig. 6. Schematic of feature extraction block
Fig. 7. Crack detection results of different algorithms. (a) Original images; (b) labelling; (c) threshold segmentation algorithm; (d) Canny algorithm; (e) proposed algorithm
Fig. 8. Visualization results with or without dataset expansion. (a) Original images; (b) labelling; (c) without dataset expansion; (d) with dataset expansion
Fig. 9. Segmentation results of proposed algorithm and mainstream algorithm. (a) Original images; (b) labelling; (c) U-Net++; (D) DeepLab-V3+; (e) PSPNet; (f) proposed algorithm
Number of training samples | With or without dataset expansion | P /% | R /% |
---|
1800 | Without dataset expansion | 48.7 | 34.1 | 8000 | With dataset expansion | 93.8 | 93.5 |
|
Table 1. Effect of HRBCS model with or without dataset expansion
Algorithm | Precision /% | Recall /% | F1 score /% | mIoU /% | Parameter /106 | |
---|
U-Net++ | 71.0 | 70.3 | 70.7 | 75.30 | 59.5 | | DeepLab-V3+ | 82.6 | 80.9 | 81.7 | 78.60 | 43.5 | | PSPNet | 95.1 | 56.9 | 71.2 | 76.00 | 65.9 | | Proposed algorithm | 93.8 | 93.5 | 93.6 | 85.48 | 65.9 | |
|
Table 2. Comparison between proposed algorithm and mainstream semantic segmentation algorithms