Author Affiliations
School of Information Engineering, East China JiaoTong University, Nanchang, Jiangxi 330013, Chinashow less
Fig. 1. Network structure of YOLOv4
Fig. 2. Schematic diagram of CSPResNet. (a) Resblock_body; (b) CSPResNet(X)t
Fig. 3. Standard convolution process
Fig. 4. Depth separable convolution process
Fig. 5. Standard convolution and depth separable convolution model based on YOLOv4 network structure. (a) Standard convolution model; (b) depth separate convolution model
Fig. 6. Examples of data set and label for road surface disease. (a) Examples of data set for road surface disease; (b) examples of label for road surface disease
Fig. 7. Examples of road surface disease data augmenting
Fig. 8. Changes of training losses based on improved YOLOv4
Fig. 9. Comparison of P-R curves for road surface disease based on different network models
Fig. 10. Detection results of road surface disease based on different network models. (a) Input images; (b) Faster R-CNN; (c) SSD; (d) YOLOv3; (e) YOLOv4; (f) YOLOv4+DC; (g) YOLOv4+FL; (h) YOLOv4+DC+FL
Name of sample | Method |
---|
Flip | Crop | Bright | Add noise |
---|
tran_cr | -- | √ | √ | √ | long_cr | -- | √ | √ | √ | mesh_cr | √ | √ | √ | √ | pothole | √ | √ | √ | √ |
|
Table 1. Data augmenting methods of different road surface disease types
Name of sample | Traning samples | Validation samples | Test samples | Total |
---|
tran_cr | 1776 | 222 | 222 | 2220 | long_cr | 1552 | 194 | 194 | 1940 | mesh_cr | 792 | 99 | 99 | 990 | pothole | 500 | 62 | 62 | 624 | Total | 4620 | 577 | 577 | 5774 |
|
Table 2. Sample distribution of road surface disease data
Name of model | Evaluation index /% |
---|
mAP | AP (tran_cr) | AP (long_cr) | AP (mesh_cr) | AP (pothole) | Time /ms |
---|
Faster R-CNN | 93.55 | 96.29 | 96.41 | 91.54 | 89.96 | 105.0 | SSD | 82.44 | 86.02 | 87.83 | 88.07 | 67.81 | 44.3 | YOLOv3 | 84.18 | 91.44 | 87.35 | 88.26 | 69.69 | 39.5 | YOLOv4 | 90.39 | 93.24 | 93.14 | 88.49 | 86.67 | 43.7 | YOLOv4+DC | 91.01 | 93.26 | 93.65 | 89.18 | 87.95 | 35.6 | YOLOv4+FL | 92.22 | 95.96 | 94.67 | 90.05 | 88.18 | 43.8 | YOLOv4+DC+FL | 93.64 | 96.37 | 96.38 | 91.77 | 90.04 | 35.8 |
|
Table 3. Comparison of detection results of road surface disease based on different network models