Optimization of Image Semantic Segmentation Algorithms Based on Deeplab v3+

Fig. 6. Influence of hetero-receptive field splicing on sampling points. (a) Sampling point distribution of convolution with rate of 12; (b) sampling point distribution of cascaded convolution with rates of (6 12)

Download full size

Fig. 7. Feature alignment module based on attention mechanism

Download full size

Fig. 8. Comparison of segmentation results on Cityscapes validation set. (a) Input images; (b) label images; (c) segmentation results of Deeplab v3+; (d) segmentation results of N-Deeplab v3+

Download full size

Fig. 9. Comparison of segmentation results on PASCAL VOC 2012 verification set. (a) Input images; (b) label images; (c) segmentation results of Deeplab v3+; (d) segmentation results of N-Deeplab v3+

Download full size

Table 1. Influence of heterogeneous field splicing

View table

Table 1. Influence of heterogeneous field splicing

Dilation rate	Receptive field	Effective operation pixel	Information utilization /%
6	13×13	3×3	5.326
12	25×25	3×3	1.440
18	37×37	3×3	0.657
6+12	37×37	7×7	3.579
6+18	49×49	9×9	3.374
12+18	61×61	9×9	2.177
6+12+18	73×73	13×13	3.171

Table 2. Performance comparison of different models on the Cityscapes dataset
View table
Table 2. Performance comparison of different models on the Cityscapes dataset
Model Backbone network mIoU /%
FCN-8S VGG-16 62.21
SegNet VGG-16 62.64
Deeplab v2 ResNet101 68.52
PSPNet ResNet101 73.98
Deeplab v3+ Xception 74.62
N-Deeplab v3+ Xception 76.31

Table 3. Quantifying information comparison of Deeplab v3+ before and after improvement
View table
Table 3. Quantifying information comparison of Deeplab v3+ before and after improvement
Model Number of parameters /10⁶ Model size /Mbit T₀ /ms Speed /（frame·s^-1）
Deeplab v3+ 43.51 165.62 275.3 3.632
N-Deeplab v3+ 37.38 142.28 302.7 3.466

Table 4. Comparison of test results of improved scheme in ASPP module

View table

Table 4. Comparison of test results of improved scheme in ASPP module

Group	Dilation rate	HFS	DSAConv	mIoU /%	Train time /h	T₀ /ms
1	（6 12 18）			74.62	23.82	275.3
2	（6 12 18 24）			74.91	25.64	310.8
3	（6 12 18）	√		75.39	27.27	322.4
4	（6 12 18 24）	√		75.71	30.44	372.0
5	（6 12 18）	√	√	75.36	21.59	253.2
6	（6 12 18 24）	√	√	75.62	25.60	312.5

Table 5. Performance comparison of different improvement schemes on the Cityscapes dataset
View table
Table 5. Performance comparison of different improvement schemes on the Cityscapes dataset
Group HFS-ASPP MFF A-FAM mIoU /% Speed /（frame·s^-1）
7 74.62 3.632
8 √ 75.36 3.949
9 √ √ 75.69 3.832
10 √ √ √ 76.31 3.466

Table 6. Performance comparison of Deeplab v3+ before and after improvement on the PASCAL VOC 2012 dataset
View table
Table 6. Performance comparison of Deeplab v3+ before and after improvement on the PASCAL VOC 2012 dataset
Method mIoU /% T₀ /ms Speed /（frame·s^-1）
Deeplab v3+ 79.83 46.86 21.340
N-Deeplab v3+ 81.97 48.35 20.612

Tools

Get Citation

Copy Citation Text

Junxi Meng, Li Zhang, Yang Cao, Letian Zhang, Qian Song. Optimization of Image Semantic Segmentation Algorithms Based on Deeplab v3+[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1610009

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: Jun. 7, 2021

Accepted: Jul. 9, 2021

Published Online: Jul. 22, 2022

The Author Email: Li Zhang (dx_zhangli@126.com)

DOI:10.3788/LOP202259.1610009

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Influence of heterogeneous field splicing

Table 1. Influence of heterogeneous field splicing

Table 2. Performance comparison of different models on the Cityscapes dataset

Table 2. Performance comparison of different models on the Cityscapes dataset

Table 3. Quantifying information comparison of Deeplab v3+ before and after improvement

Table 3. Quantifying information comparison of Deeplab v3+ before and after improvement

Table 4. Comparison of test results of improved scheme in ASPP module

Table 4. Comparison of test results of improved scheme in ASPP module

Table 5. Performance comparison of different improvement schemes on the Cityscapes dataset

Table 5. Performance comparison of different improvement schemes on the Cityscapes dataset

Table 6. Performance comparison of Deeplab v3+ before and after improvement on the PASCAL VOC 2012 dataset

Table 6. Performance comparison of Deeplab v3+ before and after improvement on the PASCAL VOC 2012 dataset