Semantic Segmentation Method Based on Multiscale Feature Alignment and Aggregation

Zhaozhong Xu; Li Peng; Feifei Dai

doi:10.3788/LOP212814

Laser & Optoelectronics Progress, Volume. 60, Issue 2, 0215004(2023)

Semantic Segmentation Method Based on Multiscale Feature Alignment and Aggregation

Zhaozhong Xu¹, Li Peng^1,2、*, and Feifei Dai³

¹Engineering Research Center of Internet of Things Technology Applications, School of IoT Engineering, Jiangnan University, Wuxi 214122, Jiangsu, China

²Jiangsu Province Internet of Things Application Technology Key Construction Laboratory, Wuxi Taihu College, Wuxi 214122, Jiangsu, China

³Taizhou Product Quality and Safety Monitoring Institute, Taizhou 318000, Zhejiang, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Structure diagram of proposed model

Download full size

Fig. 2. FAM structure diagram

Download full size

Fig. 3. SA module structure diagram

Download full size

Fig. 4. Dilated convolution

Download full size

Fig. 5. Atrous spatial pyramid pooling

Download full size

Fig. 6. Mish activation function

Download full size

Fig. 7. Visualization results of different segmentation methods. (a) Input images; (b) labels; (c) segmentation results of FCN;(d) segmentation results of Deeplab v3; (e) segmentation results of PSPNet; (f) segmentation results of proposed algorithm

Download full size

Table 1. mIoU with different feature selection
View table
Table 1. mIoU with different feature selection
Block selection Add Concat MFAA
1，2 77.5 77.5 77.6
1，3 77.7 77.6 78.0
1，4 77.5 77.3 77.9
1，2，3 77.6 77.8 78.4
1，2，4 77.5 77.7 78.4

Table 2. Experimental results of MFAA module disassembly
View table
Table 2. Experimental results of MFAA module disassembly
Module selection mIoU /%
Baseline 77.2
SA 77.7
FAM 77.8
MFAA（SA+FAM） 78.4

Table 3. Ablation results
View table
Table 3. Ablation results
Group mIoU /%
Baseline 63.8
MFAA 66.7
ASPP 76.9
ASPP+Mish 77.2
ASPP+MFAA 78.1
ASPP+MFAA+Mish 78.4

Table 4. Results of comparative experiments
View table
Table 4. Results of comparative experiments
Algorithm FCN SegNet Deeplab v3 PSPNet Proposed algorithm
mIoU /% 69.7 71.6 77.2 78.0 78.4

Tools

Get Citation

Copy Citation Text

Zhaozhong Xu, Li Peng, Feifei Dai. Semantic Segmentation Method Based on Multiscale Feature Alignment and Aggregation[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0215004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites