计算机技术与发展2025，Vol.35Issue(11)：123-129,7.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0156
基于上下文信息适配的实时语义分割网络
A Real-time Semantic Segmentation Network Based on Contextual Information Adaptation
摘要
Abstract
Most real-time semantic segmentation models currently struggle to balance efficiency and accuracy.In context modeling,global pooling is often used to extract context information at different scales.However,indiscriminate processing of different features may dilute key region features or cause background information to interfere with foreground information.To address these issues,a real-time semantic segmentation network based on Context Information Adaptation(PCMNet)is proposed,built upon the three-branch network(PIDNet).Firstly,a Dual Path Convolution Bottleneck Block(DPC-B)is designed to improve the model's expressive power while maintaining lightweight characteristics for efficient feature extraction.Secondly,the Context Information Adaptation Module(CIAM)is introduced,utilizing a self-attention mechanism to adaptively adjust context information at different spatial locations,providing more accurate segmentation results in finer details.Experimental results show that PCMNet achieves 78.9%accuracy at 109 fps on the Cityscapes validation set,and 76.3%mIoU at 133 fps on the Camvid dataset.Compared with baseline models,accuracy is improved by 0.89%and 0.79%,respectively,effectively demonstrating PCMNet's significant advantage in real-time scenarios.关键词
实时语义分割/双路径卷积/特征提取/上下文信息/自注意力机制Key words
real-time semantic segmentation/dual-path convolution/feature extraction/contextual information/self-attention mechanism分类
计算机与自动化引用本文复制引用
彭彦蓉,张智..基于上下文信息适配的实时语义分割网络[J].计算机技术与发展,2025,35(11):123-129,7.基金项目
国家重点研究计划重点专项(2022YFC3300800) （2022YFC3300800）
湖北省教育厅科学研究计划重点项目(D20231104) （D20231104）
