基于双分支多尺度特征融合的道路场景语义分割
Semantic segmentation of road scene based on dual-branch multi-scale feature fusion
肖哲璇 1陈辉 1王硕1
作者信息
- 1. 安徽理工大学计算机科学与工程学院,安徽淮南 232001
- 折叠
摘要
针对实时场景图像语义分割网络模型难以在分割精度、模型参数量和推理速度中取得平衡的问题,提出了一种轻量级实时语义分割算法.首先,该算法以双分支结构作为该网络的基本结构,并在各个分支中分别对不同分辨率的特征图进行特征信息提取;其次,在高分辨率分支和低分辨率分支中分别加入改进的金字塔切分注意力模块和残差空洞金字塔模块,并在不同分支之间进行了双侧特征融合,以充分融合空间信息和语义信息;最后,设计了特征融合模块,并通过上采样操作对图像进行恢复,以实现图像语义分割.该算法在Cityscapes和Camvid数据集上进行验证,以5.02 M的参数量分别取得了 76.8％和70.5％的mIoU,分别达到了 56 fps和147 fps的运行速度.
Abstract
Due to the imbalance between segmentation accuracy,model parameters,and inference speed in real-time scene semantic segmentation network models for scene images,a lightweight real-time semantic segmentation algorithm is proposed in this paper.Firstly,the algorithm employs a two-branch structure as the basic framework of the network,ex-tracting the feature information from different resolution feature maps in each branch.Secondly,improved pyramid split at-tention modules and residual atrous pyramid modules are incorporated into the high-resolution and low-resolution branches,respectively.Bilateral feature fusion is carried out between different branches to fully integrate spatial and semantic infor-mation.Finally,a feature fusion module is designed,and image semantic segmentation is achieved by restoring the image through upsampling operations.The algorithm achieves 76.8％and 70.5％mIoU with 5.02 M parameters and reaches up to 56 fps or 147 fps inference speed respectively in Cityscapes and Camvid datasets.
关键词
图像处理/实时语义分割/轻量级/注意力模块/多尺度特征Key words
Image processing/Real-time semantic segmentation/Lightweight/Attention module/Multi-scale features引用本文复制引用
基金项目
安徽省重点教学研究项目(2020jyxm0458)
出版年
2024
NETL
NSTL
维普
万方数据