Semantic Segmentation for Road Scene Based on Multiscale Feature Fusion

Qingming Yi; Wenting Zhang; Min Shi; Jialin Shen; Aiwen Luo

doi:10.3788/LOP220914

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 12 >Page 1210006 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 12, 1210006 (2023)

Semantic Segmentation for Road Scene Based on Multiscale Feature Fusion

Qingming Yi, Wenting Zhang, Min Shi, Jialin Shen, and Aiwen Luo^*

Author Affiliations

College of Information Science and Technology, Jinan University, Guangzhou 510632, Guangdong, China

show less

DOI: 10.3788/LOP220914 Cite this Article Set citation alerts

Qingming Yi, Wenting Zhang, Min Shi, Jialin Shen, Aiwen Luo. Semantic Segmentation for Road Scene Based on Multiscale Feature Fusion[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1210006 Copy Citation Text

show less

Fig. 1. Overall structure of MIFNet composed of three key modules marked by dashed boxes

Download full size

Fig. 2. Comparison of different bottleneck modules. (a) Non-bottleneck-1D; (b) DAB module; (c) SS-nbt module

Download full size

Fig. 3. Structure of LEF-B

Download full size

Fig. 4. Heat maps after different processing in fusion module. (a) Input image; (b) heat map without Laplace operator; (c) heat map with Laplace operator

Download full size

Fig. 5. Semantic segmentation results on Cityscapes dataset

Download full size

Network	Speed /（frame·s^-1）		Parameters /M		mIoU /%		GFLOPs
Network	Original	LFE-B	Original	LFE-B	Original	LFE-B	Original	LFE-B
DABNet^［5］	106.00	104.41	0.76	0.69	69.1	70.46	11.18	9.61
ERFNet^［7］	58.57	41.01	2.07	0.75	70.0	71.49	26.86	9.84
LEDNet^［8］	58.94	72.01	0.95	0.60	70.6	70.00	11.51	7.65
ESNet^［19］	51.39	51.72	1.66	1.38	70.7	71.12	24.35	14.29
MIFNet（proposed）	—	73.68	—	0.82	—	72.50	—	12.03

Table 1. Results of different networks with LFE-B

Configuration	Speed /（frame·s^-1）	Parameters /M	mIoU /%
None	74.97	0.82	71.71
3×3	71.80	0.83	71.65
Prewitt	71.39	0.82	71.79
Sobel	72.74	0.82	72.16
Laplace	73.68	0.82	72.50

Table 2. Results of different configuration of ESF

Decoder

Speed /（frame·s^-1）

Parameters /M

mIoU /%

None

ERFD^［7］

PAD^［5］

APN^［8］

MAFD（proposed）

88.20

52.91

75.23

67.29

73.68

0.77

1.03

0.77

0.78

0.82

71.10

72.10

71.59

69.51

72.50

Table 3. Results of different decoders on MIFNet

Network	Pretrain	Speed /（frame·s^-1）	Parameters /M	mIoU（test）/%	GFLOPs
ENet^［6］	No	41.70	0.36	58.3	4.35
ESPNet^［22］	No	146.00	0.36	60.3	3.50
CGNet^［23］	No	44.70	0.50	65.6	7.00
ContextNet^［10］	No	176.60	0.88	65.5	1.78
EDANet^［24］	No	105.50	0.68	67.3	9.00
ERFNet^［7］	No	58.57	2.07	68.0	26.90
FastSCNN^［9］	No	198.41	1.10	62.8	1.76
LEDNet^［8］	No	58.94	0.95	69.2	11.50
DABNet^［5］	No	106.20	0.64	71.2	10.50
ESNet^［19］	No	51.39	1.66	70.7	24.40
LRNNet_C^［14］	No	71.00	0.68	72.2	8.58
BiSeNetV1_X^［20］*	ImageNet	105.80^*	5.80	68.4	14.90
BiSeNetV1_R^［20］*	ImageNet	65.50^*	49.00	74.7	55.30
BiSeNetV2^［21］	No	156.00	—	72.6	21.15
BiSeNetV2_L^［21］	No	47.30	—	75.3	118.51
MIFNet（proposed）	No	73.68	0.82	73.1	12.03

Table 4. Performance comparison of different network models on Cityscapes test set

Network	Input size /pixel	Speed /（frame·s^-1）	Parameters /M	mIoU（test）/%	GFLOPs
ENet^［6］	360×480	61.00	0.36	51.3	1.44
ERFNet^［7］	360×480	64.30	2.07	67.1	8.80
DABNet^［5］	360×480	117.00	0.64	64.6	3.20
LEDNet^［8］	360×480	58.94	0.95	66.6	11.50
EKENet^［13］	360×480	38.00	1.20	67.5	—
ESPNet^［22］	360×480	132.00	0.36	55.6	1.10
EDANet^［24］	360×480	163.00	0.68	66.4	2.90
CGNet^［23］	360×480	112.00	0.50	65.6	65.60
LRNNet_C^［14］	360×480	76.50	0.68	69.2	—
BiSeNetV1_X^［20］*	720×960	175.00^*	49.00	65.6	8.70
BiSeNetV1_R^［20］*	720×960	116.30^*	5.80	68.7	32.40
BiSeNetV2^［21］	720×960	124.50	—	72.4	21.15
BiSeNetV2_L^［21］	720×960	32.70	—	73.2	118.51
MIFNet（proposed）	720×960	55.02	0.81	71.1	15.86
MIFNet（proposed）	360×480	85.16	0.81	67.7	3.90

Table 5. Performance comparison of different network models on CamVid test set

Qingming Yi, Wenting Zhang, Min Shi, Jialin Shen, Aiwen Luo. Semantic Segmentation for Road Scene Based on Multiscale Feature Fusion[J]. Laser & Optoelectronics Progress, 2023, 60(12): 1210006

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享