Lightweight Semantic Segmentation Network Based on Attention Coding

Xiaolong Chen; Ji Zhao; Siyi Chen

doi:10.3788/LOP202158.1410012

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 14 >Page 1410012 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 14, 1410012 (2021)

Lightweight Semantic Segmentation Network Based on Attention Coding

Xiaolong Chen^1、*, Ji Zhao^1、2, and Siyi Chen^1、**

Author Affiliations

¹School of Automation and Electronic Information, Xiangtan University, Xiangtan, Hunan 411100, China

²National CIMS Engineering Technology Research Center, Tsinghua University, Beijing 100084,China

show less

DOI: 10.3788/LOP202158.1410012 Cite this Article Set citation alerts

Xiaolong Chen, Ji Zhao, Siyi Chen. Lightweight Semantic Segmentation Network Based on Attention Coding[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410012 Copy Citation Text

show less

Fig. 1. Lightweight network based on attention coding

Download full size

Fig. 2. Structure of the APAM

Download full size

Fig. 3. Structure of the GAU module

Download full size

Fig. 4. Segmentation results of different algorithms. (a) Original image; (b) real semantic label; (c) basic algorithm; (d) our algorithm

Download full size

Algorithm	Base size	mIoU /%
Baseline	--	69.8
Ours-1	36	83.3
Ours-2	49	83.4
Ours-3	64	83.5
Ours-4	81	83.6
Ours-5	100	83.6

Table 1. Influence of APAM basis set number on segmentation accuracy

Algorithm	Base size	mIoU /%
MHA^[23]	--	59.95
SGE^[24]	--	58.90
PAM	--	70.71
CAM	--	70.53
Ours(average)	64	70.99
Ours(max)	64	69.93

Table 2. Segmentation results of different attention modules

CF	LCF	mIoU /%
F1	concat	84.6
F2	concat	83.2
F3	concat	83.3
F4	concat	81.4
F4	sum	84.5
F5	sum	84.9
F6	sum	84.9

Table 3. Structural analysis of the APAM

Algorithm	GFLOPS	Params /M	Memory /G	mIoU /%
FCN	216.0	54.0	7.7	69.8
U-net	262.2	34.5	8.7	70.8
SegNet	244.6	32.4	8.9	71.1
DeepLab v2	251.7	44.6	8.0	71.6
PSP	254.7	67.6	8.1	80.9
DANet	275.4	68.5	9.3	85.1
Ours	228.8	67.7	8.1	84.9

Table 4. Performance parameters of different algorithms

Algorithm	FCN	DeepLab v2	DPN^[25]	DeepLab v3+	PSP	ResNet38^[26]	EncNet^[27]	Ours
aero	82.4	84.4	87.7	88.0	87.4	94.4	94.1	91.6
bike	47.4	54.5	59.4	56.3	56.3	72.9	69.2	57.9
bird	81.2	81.5	78.4	86.3	85.7	94.9	96.3	90.0
boat	68.6	63.6	64.9	69.4	79.4	68.8	76.7	85.5
bottle	75.3	65.9	70.3	72.2	73.8	78.4	86.2	82.5
bus	81.3	85.1	89.3	90.3	92.3	90.6	96.3	95.0
car	79.9	79.1	83.5	85.7	87.3	90.0	90.7	90.8
cat	81.6	83.4	86.1	89.6	92.3	92.1	94.2	94.0
chair	33.7	30.7	31.7	28.9	53.3	40.1	38.8	53.8
cow	68.4	74.1	79.9	85.9	90.4	90.4	90.7	93.7
table	52.3	59.8	62.6	59.3	75.2	71.7	73.3	78.0
dog	76.4	79.0	81.9	84.2	87.3	89.9	90.0	93.0
horse	64.9	76.1	80.0	80.2	85.9	93.7	92.5	91.4
mbike	73.4	83.2	83.5	84.2	83.8	91.0	88.8	87.1
person	81.2	80.8	82.3	82.8	84.5	89.1	87.9	88.2
plant	56.7	59.7	60.5	56.0	68.1	71.3	68.7	74.1
sheep	69.7	82.2	83.2	78.5	87.0	90.7	92.6	91.3
sofa	50.9	50.4	53.4	51.6	73.0	61.3	59.0	75.8
train	78.5	73.1	77.9	84.5	91.1	87.7	86.4	92.8
tv	70.1	63.7	65.0	69.6	71.5	78.1	73.4	80.6
mIoU	69.8	71.6	74.1	75.1	80.9	82.5	82.9	84.9

Table 5. Classification results of different algorithms on the PASCAL VOC2012 validation set unit: %

Algorithm	FCN	PSP	DenseASPP^[28]	DANet	Ours
road	95.1	96.4	97.3	97.2	97.4
sidewalk	67.8	74.4	78.1	77.8	79.4
building	88.5	89.1	89.5	89.8	90.0
wall	50.5	52.9	56.0	56.1	57.0
fence	44.6	47.9	48.5	48.6	51.1
pole	35.6	39.9	40.3	40.8	43.5
traffic light	47.0	51.9	52.8	53.0	53.4
Traffic sign	60.4	62.4	65.3	65.2	66.5
vegetation	88.6	89.4	89.6	89.7	89.8
terrain	55.6	57.6	60.5	60.7	60.9
sky	91.4	92.0	92.3	92.4	92.7
person	68.8	70.4	71.4	71.9	72.8
rider	47.9	49.9	52.0	52.2	53.4
car	90.3	91.4	92.1	92.4	92.4
truck	73.8	73.9	83.0	82.8	79.3
bus	73.6	75.8	80.2	79.4	81.9
train	62.8	66.4	70.4	70.8	74.3
motocycle	51.7	55.0	58.4	58.9	58.6
bicycle	63.1	63.6	65.5	65.8	66.6
mIoU	66.2	68.4	70.7	70.8	71.6

Table 6. Classification results of different algorithms on the Cityscapes validation set unit: %

Xiaolong Chen, Ji Zhao, Siyi Chen. Lightweight Semantic Segmentation Network Based on Attention Coding[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410012

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information