.png){kind=icon}
.png){kind=icon}
Return
Multi-Path Feature Enhancement Network for real-time semantic segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.130099.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Convolutional Neural Networks (CNN) have been widely used in semantic segmentation. But it mainly focuses on extracting the local hierarchical information, and the ability of the global information extraction is insufficient. Transformer is good at extracting the global information, but it is computationally intensive and unfriendly to real-time semantic segmentation. In this work, we propose a Multi-Path Feature Enhancement Network (MPFENet) for real-time semantic segmentation. Its backbone is CNN and uses the Multi-Path Feature Enhancement (MPFE) module to enhance the global feature information by Transformer. We propose a Dual Pooling Transformer (DPFormer) to extract the salient and global information while reducing the computational effort. We further propose a Parallel Aggregation Asymmetric Pyramid Pooling Module (PAAPPM) to extract different scale and orientation information. Finally, we propose a Multi-Scale Feature Fusion Module (MSFFM) to fuse the features of different paths processed by feature enhancement. Numerous experiments on five datasets demonstrate that our MPFENet is effective and achieves satisfactory results. Specifically, MPFENet achieves 80.2% mIoU at 113.8 FPS on the Cityscapes dataset and 79.7% mIoU at 151.2 FPS on the CamVid dataset.
Keywords:
Semantic segmentation
Feature enhancement
Dual pooling transformer
Real-time