Dual-Branch Remote Sensing Building Extraction Network Based on Texture Enhancement

Fig. 4. Visualization results of ablation experiments。(a) Images; (b) labels; (c) Baseline; (d) Baseline+Out; (e) Baseline+Out+EAM; (f) proposed model

Download full size

Fig. 5. Visualized results on the Massachusetts Building dataset. (a) Images; (b) labels; (c) TEOA-UNet; (d) SDSC-UNet; (e) MANet; (f) BANet; (g) DC-Swin

Download full size

Fig. 6. Visualized results on the Massachusetts Building dataset. (a) Images; (b) labels; (c) TEOA-UNet; (d) SDSC-UNet; (e) MANet; (f) BANet; (g) DC-Swin

Download full size

Fig. 7. Visualized results on the Inria dataset. (a) Images; (b) labels; (c) TEOA-UNet; (d) SDSC-UNet; (e) MANet; (f) BANet; (g) DC-Swin

Download full size

Table 1. Texture enhancement branch
View table
Table 1. Texture enhancement branch
Layer Input size Output size Component k c s p
#01 3×H×W 64× $\frac{H}{2}$ × $\frac{W}{2}$ ConvBR 7 64 2 3
#02 64× $\frac{H}{2}$ × $\frac{W}{2}$ 64× $\frac{H}{4}$ × $\frac{W}{4}$ ConvBR 3 64 2 1
#03 64× $\frac{H}{4}$ × $\frac{W}{4}$ C_g× $\frac{H}{8}$ × $\frac{W}{8}$ ConvBR 3 C_g 2 1
#04 C_g× $\frac{H}{8}$ × $\frac{W}{8}$ 1× $\frac{H}{8}$ × $\frac{W}{8}$ ConvBR 1 1 1 0

Table 2. Results of ablation experiment
View table
Table 2. Results of ablation experiment
Model R_IoU R_precision R_recall s_F1
Baseline 76.71 88.05 85.62 86.82
Baseline+Out 77.70 87.80 87.09 87.45
Baseline+Out+EAM 79.10 89.13 87.54 88.32
Proposed model 79.45 88.97 88.12 88.54

Table 3. Comparison results of hyperparameters in the Massachusetts dataset

View table

Table 3. Comparison results of hyperparameters in the Massachusetts dataset

$a_{1}$	$a_{2}$	$a_{3}$	$a_{4}$	R_IoU	R_precision	R_recall	s_F1
0.5	1	10	0.2	88.26	78.98	88.28	88.07
2	1	10	0.2	88.01	78.58	88.61	87.41
1	0.5	10	0.2	87.79	77.15	88.29	87.45
1	2	10	0.2	88.31	78.49	87.71	87.47
1	1	5	0.2	88.03	78.27	87.65	87.23
1	1	15	0.2	87.98	78.54	88.24	87.72
1	1	10	0.1	88.09	78.71	89.06	87.14
1	1	10	0.5	88.27	79.01	88.35	88.03
1	1	10	0.2	88.54	79.45	88.97	88.12

Table 4. Accuracy comparison of Massachusetts Building dataset
View table
Table 4. Accuracy comparison of Massachusetts Building dataset
Model R_IoU R_precision R_recall s_F1
MANet^［17］ 70.76 82.00 83.77 82.88
BANet^［18］ 72.20 83.07 84.66 83.86
DC-Swin^［22］ 72.59 83.07 85.19 84.12
SDSC-UNet^［24］ 76.71 88.05 85.62 86.82
TEOA-UNet 79.45 88.97 88.12 88.54

Table 5. Accuracy comparison of WHU Building dataset
View table
Table 5. Accuracy comparison of WHU Building dataset
Model R_IoU R_precision R_recall s_F1
MANet^［17］ 90.04 94.61 94.89 94.75
BANet^［18］ 89.09 94.54 94.89 94.75
DC-Swin^［22］ 89.10 94.43 94.04 94.23
SDSC-UNet^［24］ 90.21 94.32 95.15 94.73
TEOA-UNet 91.88 95.52 95.18 95.22

Table 6. Accuracy comparison of Inria dataset
View table
Table 6. Accuracy comparison of Inria dataset
Model R_IoU R_precision R_recall s_F1
MANet^［17］ 76.22 86.54 86.47 86.52
BANet^［18］ 76.9 87.73 86.17 86.94
DC-Swin^［22］ 78.03 88.62 87.48 88.04
SDSC-UNet^［24］ 83.01 91.52 89.92 90.71
TEOA-UNet 82.63 91.67 90.23 90.94

Table 7. Cross domain experimental results
View table
Table 7. Cross domain experimental results
Dataset Model R_IoU R_precision R_recall s_F1
Massachusetts Baseline 41.47 83.20 44.91 58.63
Proposed model 43.75 84.42 47.99 60.87
Inria Baseline 58.97 82.47 67.42 74.19
Proposed model 61.72 82.49 71.79 76.33

Tools

Get Citation

Copy Citation Text

Xu Chen, Mingchang Shi. Dual-Branch Remote Sensing Building Extraction Network Based on Texture Enhancement[J]. Laser & Optoelectronics Progress, 2024, 61(14): 1428006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites