Fig. 1. Comparison between RSANet and LinkNet. (a) RSANet; (b) LinkNet
Fig. 2. Structure of encoder block
Fig. 3. Structure of decoder block
Fig. 4. Position attention module
Fig. 5. Channel attention module
Fig. 6. Remote sensing images and road labels in Massachusetts Roads dataset. (a) Part of the remote sensing images in the training set; (b) road labels corresponding to images in the training set; (c) some remote sensing images in the test set; (d) road labels corresponding to images in the test set
Fig. 7. Remote sensing images and road labels in DeepGlobe dataset. (a) Part of the remote sensing images in the training set; (b) road labels corresponding to images in the training set; (c) some remote sensing images in the test set; (d) road labels corresponding to images in the test set
Fig. 8. Segmentation results of different depth models in the Massachusetts Roads test set. (a) Original images; (b) labels; (c) segmentation results of LinkNet; (d) segmentation results of RSANet
Fig. 9. Segmentation results of different depth models on the DeepGlobe Road Extraction test set. (a) Original images; (b) labels; (c) segmentation results of LinkNet; (d) segmentation results of RSANet
Model | Image 1 | Image 2 | Image 3 | Image 4 | Image 5 | | | | |
---|
Pr | F1-score | Pr | F1-score | | | | Pr | F1-score | Pr | F1-score | Pr | F1-score |
---|
LinkNet[3] | 0.839 | 0.859 | 0.913 | 0.881 | 0.821 | 0.832 | 0.800 | 0.817 | 0.747 | 0.779 | RSANet | 0.979 | 0.976 | 0.923 | 0.892 | 0.829 | 0.835 | 0.822 | 0.889 | 0.818 | 0.827 |
|
Table 1. Segmentation results of depth models on each test image in Fig. 8
Model | Pr | F1-score | Training time /h | Inference time per picture /s |
---|
LinkNet[3](baseline) | 0.802 | 0.815 | 31 | 0.31 | LinkNet-PAM | 0.812 | 0.823 | 32 | 0.32 | LinkNet-CAM | 0.818 | 0.828 | 40 | 0.40 | RSANet | 0.827 | 0.843 | 41 | 0.41 |
|
Table 2. Segmentation results of different depth models on the Massachusetts Roads test set
Model | Image 6 | Image 7 | Image 8 | Image 9 | Image 10 | | | | |
---|
Pr | PIOU | Pr | PIOU | | | | Pr | PIOU | Pr | PIOU | Pr | PIOU |
---|
LinkNet[3] | 0.770 | 0.712 | 0.597 | 0.515 | 0.733 | 0.631 | 0.886 | 0.722 | 0.729 | 0.704 | RSANet | 0.792 | 0.735 | 0.824 | 0.696 | 0.835 | 0.694 | 0.907 | 0.790 | 0.773 | 0.723 |
|
Table 3. Segmentation results of different depth models on each test image in Fig. 9
Model | PP | PIOU | Training time /h | Inference time per picture /s |
---|
LinkNet[3](baseline) | 0.785 | 0.598 | 31 | 0.31 | LinkNet-PAM | 0.791 | 0.610 | 32 | 0.32 | LinkNet-CAM | 0.796 | 0.616 | 40 | 0.40 | RSANet | 0.811 | 0.624 | 41 | 0.41 |
|
Table 4. Segmentation results of different depth models on the DeepGlobe Road Extraction test set
Model | Pr | PP | F1-score |
---|
FCN-4s[2] | 0.660 | 0.710 | 0.684 | SegNet[18] | 0.765 | 0.773 | 0.768 | ELU-SegNet[18] | 0.773 | 0.852 | 0.788 | ELU-SegNet-R[18] | 0.780 | 0.847 | 0.812 | DCED[19] | 0.839 | 0.825 | 0.829 | RSANet | 0.827 | 0.859 | 0.843 |
|
Table 5. Segmentation results of different depth models on the Massachusetts Roads test set
Model | PP | F1-score | Training time /h | Inference time per picture /s |
---|
U-Net[1] | 0.764 | 0.597 | 70 | 0.94 | SegNet[20] | 0.774 | 0.602 | 98 | 1.8 | LinkNet[3](baseline) | 0.785 | 0.598 | 31 | 0.31 | RSANet | 0.811 | 0.624 | 41 | 0.41 |
|
Table 6. Segmentation results of different depth models on the DeepGlobe Road Extraction test set