基于改进高分辨率网络的多语义图像分割方法  

A Multi Semantic Image Segmentation Method Based on Improved High Resolution Networks

作　　者：张少杰 彭富明[1] 方斌[1] 张子祥 相福磊 何浩天 ZHANG Shaojie;PENG Fuming;FANG Bin;ZHANG Zixiang;XIANG Fulei;HE Haotian(Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学,江苏南京210094

出　　处：《机械制造与自动化》2024年第3期181-184,共4页Machine Building & Automation

基　　金：国家重点研发计划项目(2021YFE0194600);江苏省科技计划项目(BZ2023023)。

摘　　要：针对室外复杂场景下图像分割难度较大的问题,提出一种基于HRNet的多语义图像分割模型(HR_DfeNet)。该模型通过引入通道注意力和空间注意力模块优化特征提取,通过改进金字塔池化模块设计ASPP_M模块形成高分辨率特征提取分支,并与多种注意力机制融合。在Cityscape数据集上,HR_DfeNet相较于传统分割模型表现出不同程度的分割优化效果。To address the difficulty of image segmentation in complex outdoor scenes,this paper proposes a multi semantic image segmentation model based on HRNet(HR_DfeNet),which optimizes feature extraction by introducing channel attention and spatial attention modules,designs a high-resolution feature extraction branch by improving the pyramid pooling module and ASPP_M module,and integrates with multiple attention mechanisms.On the Cityscape dataset,HR_DfeNet exhibits varying degrees of segmentation optimization performance compared to traditional segmentation models.

关 键 词：室外复杂场景 图像分割 注意力模块 金字塔池化模块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]