基于三分支网络的实时图像语义分割

打开文本图片集

关键词：语义分割；深度学习；实时性；注意力机制；多尺度特征中图分类号：TP391.41 文献标识码：Adoi：10.37188/OPE.20263401.0167 CSTR：32169.14.OPE.20263401.0167

Real-time image semantic segmentation based on three-branchnetwork

REN Fenglei 1，2 ， GAO Ziyang 1，2 ， ZHANG Yan 1，2 ， ZHOU Haibo 1，2* ，YANGLu 1，2 ，QIN Zhichang 1，2

（1. Tianjin Key Laboratory for Aduanced Mechatronic System Design and Intelligent Control， Tianjin University ofTechnology， Tianjin 3OO384，China; 2.National Demonstration Center for Experimental Mechanical and Electrical Engineering Education， Tianjin University of Technology，Tianjin 3OO384，China） * Corresponding author， E -mail ： haibo_zhou@163. com

Abstract： To meet the stringent requirements for both accuracy and real-time performance in applications such as autonomous driving，a real-time image semantic segmentation algorithm based on a triple-branch network is proposed to achieve a favorable balance between segmentation accuracy and inference speed. Inspired by PIDNet，a triple-branch architecture is designed to extract fine-grained detail information，semantic contextual information，and edge cues from the input image，respectively. An efficient pyramid pooling module is integrated into the semantic context branch to capture multi-scale contextual information and enlarge the network receptive field. In adition，a lightweight multi-scale channel interaction attention mechanism is introduced into both the detail and edge branches to enhance feature representations.Features from the three branches are subsequently fused，and a semantic segmentation head is employed to produce the final result. The proposed network achieves 79.2% mIoU at 88.5 frame/s on the Cityscapes dataset and 80.5% mIoU at 14O.1 frame/s on the CamVid dataset. Experimental results demonstrate that the proposed method performs semantic segmentation efficiently，providing an efective trade-off between real-time performance and accuracy while significantly outperforming existing baseline methods.

Key words： semantic segmentation；deep learning;real time; atention mechanism； multi-scale features

1引言

语义分割作为计算机视觉领域的一项关键任务，其核心目标是对图像的所有像素进行精确分类，并赋予相应的标签[1]。（剩余16134字）