Remote Sensing Image Segmentation Method Based on Multi-Level Channel Attention

Shuai Yu; Xili Wang

doi:10.3788/LOP57.041012

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 4 >Page 041012 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 4, 041012 (2020)

Remote Sensing Image Segmentation Method Based on Multi-Level Channel Attention

Shuai Yu and Xili Wang^*

Author Affiliations

School of Computer Science, Shaanxi Normal University, Xi′an, Shaanxi 710119, China

show less

DOI: 10.3788/LOP57.041012 Cite this Article Set citation alerts

Shuai Yu, Xili Wang. Remote Sensing Image Segmentation Method Based on Multi-Level Channel Attention[J]. Laser & Optoelectronics Progress, 2020, 57(4): 041012 Copy Citation Text

show less

Fig. 1. Structure of RSIS-MLCA

Download full size

Fig. 2. Channel attention module

Download full size

Fig. 3. Partial training set (first line) and test set (second line) images in Massachusetts road data set

Download full size

Fig. 4. Partial training set (first line) and test set (second line) images in Inria aerial image labeling data set

Download full size

Fig. 5. Depth model segmentation results in Massachusetts road data set. (a) RGB images; (b) ground-truth images; (c) Unet segmentation results; (d) RSIS-MLCA segmentation results

Download full size

Fig. 6. Training process curves in Massachusetts road data set. (a) x_loss curve on training set; (b) R_IOU curve on test set

Download full size

Fig. 7. Depth model segmentation results in Inria aerial image labeling data set. (a) RGB images; (b) ground-truth images; (c) Unet segmentation results; (d) RSIS-MLCA segmentation results

Download full size

Fig. 8. Training process curves in Inria aerial image labeling data set. (a) x_loss curve on training set; (b) R_IOU curve on test set

Download full size

Operation	Parameter
Operation	Input_dimension	Out_dimension	Kernel_size	Stride	Padding	Image_size
Conv_1×2	3	64	3	1	1	256×256
AT_block_1	64	64				256×256
Maxpool_1	64	64	2	2	0	128×128
Conv_2×2	64	128	3	1	1	128×128
AT_block_2	128	128				128×128
Maxpool_2	128	128	2	2	0	64×64
Conv_3×2	128	256	3	1	1	64×64
AT_block_3	256	256				64×64
Maxpool_3	256	256	2	2	0	32×32
Conv_4×2	256	512	3	1	1	32×32
AT_block_4	512	512				32×32
Maxpool_4	512	512	2	2	0	16×16
Conv_5	512	1024	3	1	1	16×16
AT_block_5	1024	1024				16×16
Deconv_6	1024	512	2	2	0	32×32
Conv_6	1024	512	3	1	1	32×32
Deconv_7	512	256	2	2	0	64×64
Conv_7	512	256	3	1	1	64×64
Deconv_8	256	128	2	2	0	128×128
Conv_8	256	128	3	1	1	128×128
Deconv_9	128	64	2	2	0	256×256
Conv_9	128	64	3	1	1	256×256
Conv_10	64	1	1	1	0	256×256

Table 1. Network parameters of RSIS-MLCA

Model	Image1		Image2		Image3		Image4		Image5		Test set
Model	P	R_IOU	P	R_IOU	P	R_IOU	P	R_IOU	P	R_IOU	P	R_IOU
Unet	88.0	70.0	89.9	50.4	87.9	63.0	74.6	41.4	88.5	76.6	83.0	74.6
RSIS-MLCA	89.3	80.9	91.2	83.4	90.1	70.5	81.2	65.6	90.7	80.3	85.2	76.5

Table 2. Evaluation results for each picture in Fig. 5 and test set%

Model	Image1		Image2		Image3		Image4		Image5		Test set
Model	P	R_IOU	P	R_IOU	P	R_IOU	P	R_IOU	P	R_IOU	P	R_IOU
Unet	89.3	89.9	90.1	70.2	91.8	76.2	93.3	87.7	81.1	71.2	80.6	71.1
RSIS-MLCA	89.7	91.2	91.4	73.0	94.3	83.8	94.3	89.1	87.3	78.3	85.6	75.5

Table 3. Evaluation results for each picture in Fig. 7 and test set%

Method	P	R	R_F1-score
FCN-4s^[15]	71.0	66.0	68.4
FCN-no-skip^[16]	74.2	74.2	74.2
FCN-8s^[16]	76.2	76.2	76.2
SegNet^[16]	77.3	76.5	76.8
ELU-SegNet^[17]	85.2	73.3	78.8
ELU-SegNet-R^[17]	84.7	78.0	81.2
RSIS-MLCA	85.2	81.4	83.2

Table 4. Comparison between existing results and RSIS-MLCA%

Method	R_IOU/%
FCN^[18]	53.96
FCN-Skip^[18]	63.17
FCN-MLP^[18]	64.88
Mask R-CNN^[19]	59.37
RiFCN^[20]	74.49
RSIS-MLCA	75.53

Table 5. Comparison between existing results and RSIS-MLCA in Inria aerial image labeling data set

Shuai Yu, Xili Wang. Remote Sensing Image Segmentation Method Based on Multi-Level Channel Attention[J]. Laser & Optoelectronics Progress, 2020, 57(4): 041012

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information