Semantic Segmentation of Remote Sensing Imagery Based on Improved Squeeze and Excitaion Block

Shengwei Wu; Jiaoli Fang; Daming Zhu

doi:10.3788/LOP231528

Laser & Optoelectronics Progress, Volume. 61, Issue 12, 1228002(2024)

Semantic Segmentation of Remote Sensing Imagery Based on Improved Squeeze and Excitaion Block

Shengwei Wu¹, Jiaoli Fang^2、*, and Daming Zhu¹

Author Affiliations

¹Faculty of Land Resources Engineering, Kunming University of Science and Technology, Kunming 650000, Yunnan , China

²Computer Center, Kunming University of Science and Technology, Kunming 650000, Yunnan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(10)

Fig. 1. Residual block structure before and after improvement. (a) SE-ResNet; (b) RSE-ResNet

Download full size

Fig. 2. RSE-module: structure of the muti-scale RSE-block integration

Download full size

Fig. 3. Overall structure of the network. (a) RSENet-V1; (b) RSENet-V2; (c) V2 decoder_block

Download full size

Fig. 4. Visualization analysis results on the Potsdam dataset

Download full size

Fig. 5. Examples of segmentation results on the Potsdam dataset

Download full size

Table 1. Structure of RSE-ResNet

View table

Table 1. Structure of RSE-ResNet

Layer name	RSE-RseNet18	RSE-RseNet36	RSE-RseNet50
layer_1	Conv，7 ×7，64，stride 2
layer_1	maxpool，3×3，stride 2
layer_2	$[\begin{array}{l} C o n v, 3 \times 3,64 \\ C o n v, 3 \times 3,64 \\ R S E - b l o c k, [k, r] \end{array}] \times 2$	$[\begin{array}{l} C o n v, 3 \times 3,64 \\ C o n v, 3 \times 3,64 \\ R S E - b l o c k, [k, r] \end{array}] \times 3$	$[\begin{array}{l} C o n v, 3 \times 3,64 \\ C o n v, 3 \times 3,64 \\ R S E - b l o c k, [k, r] \end{array}] \times 3$
layer_3	$[\begin{array}{l} C o n v, 3 \times 3,128 \\ C o n v, 3 \times 3,128 \\ R S E - b l o c k, [k, r] \end{array}] \times 2$	$[\begin{array}{l} C o n v, 3 \times 3,128 \\ C o n v, 3 \times 3,128 \\ R S E - b l o c k, [k, r] \end{array}] \times 4$	$[\begin{array}{l} C o n v, 3 \times 3,128 \\ C o n v, 3 \times 3,128 \\ R S E - b l o c k, [k, r] \end{array}] \times 4$
layer_4	$[\begin{array}{l} C o n v, 3 \times 3,256 \\ C o n v, 3 \times 3,256 \\ R S E - b l o c k, [k, r] \end{array}] \times 2$	$[\begin{array}{l} C o n v, 3 \times 3,256 \\ C o n v, 3 \times 3,256 \\ R S E - b l o c k, [k, r] \end{array}] \times 6$	$[\begin{array}{l} C o n v, 3 \times 3,256 \\ C o n v, 3 \times 3,256 \\ R S E - b l o c k, [k, r] \end{array}] \times 6$
layer_5	$[\begin{array}{l} C o n v, 3 \times 3,512 \\ C o n v, 3 \times 3,512 \\ R S E - b l o c k, [k, r] \end{array}] \times 2$	$[\begin{array}{l} C o n v, 3 \times 3,512 \\ C o n v, 3 \times 3,512 \\ R S E - b l o c k, [k, r] \end{array}] \times 3$	$[\begin{array}{l} C o n v, 3 \times 3,512 \\ C o n v, 3 \times 3,512 \\ R S E - b l o c k, [k, r] \end{array}] \times 3$

Table 2. Influence of the value of the hyperparameter $k$ on the network
View table
Table 2. Influence of the value of the hyperparameter $k$ on the network
Network mIoU/F₁
$k = 5$ $k = 10$ $k = 15$ $k = 30$
RSE-ResNet18 0.711/0.827 0.719/0.833 0.722/0.836 0.716/0.829
RSE-ResNet34 0.714/0.829 0.721/0.834 0.724/0.836 0.722/0.833
RSE-ResNet50 0.709/0.825 0.714/0.830 0.717/0.832 0.716/0.829

Table 3. Comparison results of different networks with different depths

View table

Table 3. Comparison results of different networks with different depths

Network	F₁						mIoU	Mean F₁	Params
Network	Impervious surfaces	Building	Low vegetation	Tree	Car	Clutter	mIoU	Mean F₁	Params
ResNet18	0.864	0.887	0.820	0.784	0.872	0.685	0.697	0.819	—
SE-ResNet18	0.864	0.891	0.824	0.773	0.864	0.711	0.701	0.821	8.70×10⁴
CBAM-18	0.869	0.897	0.828	0.785	0.858	0.699	0.703	0.823	1.70×10⁵
RSE-ResNet18	0.868	0.901	0.831	0.786	0.878	0.730	0.719	0.832	9.00×10⁴
ResNet34	0.870	0.901	0.826	0.788	0.875	0.680	0.701	0.823	—
SE-ResNet34	0.869	0.903	0.827	0.787	0.879	0.710	0.715	0.829	1.57×10⁵
CBAM-34	0.875	0.909	0.832	0.780	0.873	0.735	0.720	0.834	3.14×10⁵
RSE-ResNet34	0.876	0.912	0.833	0.788	0.881	0.737	0.724	0.838	1.62×10⁵
ResNet50	0.869	0.895	0.827	0.783	0.877	0.703	0.710	0.826	—
SE-ResNet50	0.874	0.884	0.836	0.776	0.879	0.721	0.716	0.829	2.50×10⁶
CBAM-50	0.874	0.896	0.830	0.791	0.882	0.716	0.717	0.832	5.00×10⁶
RSE-ResNet50	0.871	0.905	0.830	0.782	0.883	0.748	0.724	0.836	2.60×10⁶

Table 4. Results of RSE-module effectiveness experiment
View table
Table 4. Results of RSE-module effectiveness experiment
Network mIoU F₁
Baseline 0.710 0.827
CRF 0.709 0.825
ASPP（6，12，18）^［7］ 0.717 0.831
PPM（1，3，6，9）^［6］ 0.718 0.833
RSE-module（5，10，15） 0.728 0.840

Table 5. Semantic segmentation results on the Potsdam dataset

View table

Table 5. Semantic segmentation results on the Potsdam dataset

Network	F₁						mIoU	Mean F₁
Network	Impervious surfaces	Building	Low vegetation	Tree	Car	Clutter	mIoU	Mean F₁
RSENet-V2	0.886	0.916	0.835	0.815	0.882	0.756	0.738	0.848
FCN-32s^［1］	0.864	0.882	0.811	0.802	0.855	0.689	0.694	0.817
Baseline	0.869	0.895	0.827	0.791	0.877	0.703	0.710	0.827
FCN-8s^［1］	0.875	0.900	0.833	0.799	0.851	0.710	0.711	0.828
HRNet-V2^［31］	0.880	0.913	0.810	0.805	0.877	0.679	0.709	0.828
DeepLabV3+^［29］	0.877	0.905	0.837	0.799	0.880	0.722	0.724	0.837
U-Net^［32］	0.878	0.898	0.832	0.815	0.897	0.714	0.728	0.839
RSENet-V1	0.875	0.907	0.833	0.784	0.885	0.756	0.728	0.840
SegNet^［33］	0.879	0.902	0.846	0.813	0.889	0.713	0.730	0.840

Tools

Get Citation

Copy Citation Text

Shengwei Wu, Jiaoli Fang, Daming Zhu. Semantic Segmentation of Remote Sensing Imagery Based on Improved Squeeze and Excitaion Block[J]. Laser & Optoelectronics Progress, 2024, 61(12): 1228002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites