Remote Sensing Image Segmentation Network Based on Adaptive Multiscale and Contour Gradient

Mengjia Niu; Yongjun Zhang; Zhi Li; Gang Yang; Zhongwei Cui; Junwen Liu

doi:10.3788/LOP220525

Laser & Optoelectronics Progress, Volume. 60, Issue 2, 0228009(2023)

Remote Sensing Image Segmentation Network Based on Adaptive Multiscale and Contour Gradient

Mengjia Niu¹, Yongjun Zhang^1、*, Zhi Li¹, Gang Yang², Zhongwei Cui³, and Junwen Liu¹

Author Affiliations

¹College of Computer Science and Technology, Guizhou University, Guiyang 550025, Guizhou, China

²Guiyang Orbita Aerospace Science＆Technology Co., Ltd., Guiyang 550027, Guizhou, China

³Big Data Science and Intelligent Engineering Research Institute, Guizhou Education University, Guiyang 550018, Guizhou, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Framework of remote sensing image semantic segmentation algorithm based on fused contour learning with deep convolutional neural network

Download full size

Fig. 2. Multi-channel network framework based on attention mechanism

Download full size

Fig. 3. Diagrams of internal structure of MA block. (a) Internal structure of a single MA block; (b) SK weight module

Download full size

Fig. 4. Diagram of internal structure of D-MA block

Download full size

Fig. 5. Visualization of ablation experiments compared in Vaihingen test set

Download full size

Fig. 6. Comparison of prediction results with mainstream models on Vaihingen test set

Download full size

Fig. 7. Comparison of prediction results with mainstream models on Potsdam test set

Download full size

Fig. 8. Comparison of prediction results with mainstream models on WHU building test set

Download full size

Table 1. Configuration parameters for profile extraction module

View table

Table 1. Configuration parameters for profile extraction module

Layer	Output size	Operator	Stride	Size
Branch-1	256×256，64	MA block	1	1
Branch-2	256×256，64	MA block	1	1
Down layer-1	128×128，64	Conv-block	2	1/2
Branch-3	128×128，64	MA block	1	1/2
Branch-4	128×128，64	MA block	1	1/2
Down layer-2	64×64，64	Conv-block	2	1/4
Branch-5	64×64，128	MA block	1	1/4

Table 2. Model ablation experiments on Vaihingen dataset
View table
Table 2. Model ablation experiments on Vaihingen dataset
Model mIoU /% OA /% F1 /% Params FLOPs /10⁹
SegNet 76.50 87.96 86.99 20.73 27.56
Improved SegNet 80.69 88.63 89.81 20.91 28.29
with D-MMA 84.28 89.93 90.36 21.82 31.39
with D- MMA +CME 86.56 92.93 92.51 22.69 34.36

Table 3. Comparison with other networks on Vaihingen dataset

View table

Table 3. Comparison with other networks on Vaihingen dataset

Model	IoU					F1	OA	mIoU
Model	impervious surfaces	building	low vegetation	tree	car	F1	OA	mIoU
U-Net	79.45	85.23	64.93	74.77	38.51	86.94	85.43	68.58
SegNet	81.69	86.41	73.43	78.36	42.63	86.99	87.96	76.50
ERFNet	77.51	79.27	62.35	71.27	35.29	83.57	82.09	65.13
PSPNet	87.69	91.94	81.52	84.79	55.79	89.86	90.19	81.16
DSMNet^［25］						90.82	91.5
Fres- MFDNN^［26］						92.0	91.0	85.0
Proposed model	91.89	93.61	86.72	88.69	71.93	92.51	92.93	86.56

Table 4. Comparison with other networks on Potsdam dataset

View table

Table 4. Comparison with other networks on Potsdam dataset

Model	IoU					F1	OA	mIoU
Model	impervious surfaces	building	low vegetation	tree	car	F1	OA	mIoU
U-Net	76.44	83.22	65.92	58.03	71.87	83.12	77.45	71.10
SegNet	83.63	91.72	74.70	71.67	78.00	85.31	87.45	77.94
ERFNet	61.46	74.78	51.83	45.85	17.07	80.57	72.35	50.20
PSPNet	83.59	92.99	76.28	73.09	77.11	85.78	89.78	80.61
BAM-Unet-sc^［27］						88.59	89.13
ResUNet-a^［3］						92.09	91.50
Proposed model	85.39	93.64	78.82	76.48	81.57	91.65	92.18	83.18

Table 5. Comparison with other network indicators on WHU building dataset
View table
Table 5. Comparison with other network indicators on WHU building dataset
Model F1 OA mIoU
U-Net 89.33 87.59 83.01
SegNet 93.73 92.53 88.76
ERFNe 89.33 87.59 78.72
PSPNet 93.28 92.15 87.26
DeNet^［28］ 94.80 90.12
MA-FCN^［29］ 95.15 90.70
Proposed model 95.81 94.06 92.30

Tools

Get Citation

Copy Citation Text

Mengjia Niu, Yongjun Zhang, Zhi Li, Gang Yang, Zhongwei Cui, Junwen Liu. Remote Sensing Image Segmentation Network Based on Adaptive Multiscale and Contour Gradient[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0228009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites