全局感知与多尺度特征融合的城市道路语义分割

打开文本图片集

关键词：深度学习；图像处理；语义分割；特征融合；损失函数中图分类号：TP391.4 文献标识码：Adoi：10.37188/OPE.20253314.2262 CSTR：32169.14.OPE.20253314.2262

Urban road semantic segmentation with global awareness and multi-scale feature fusion

WUKaijun1，ZHANG Zhiruil*，WANGYing1，ANLiwei²

（1.School of Electronic and Information Engineering， Lanzhou Jiaotong University， Lanzhou 730070 ，China;2.Pratacultural College， Inner Mongolia Minzu University，Tongliao O28OOO，China） * Corresponding author，E-mail：zzr_ljd@l63.com

Abstract： Semantic segmentation plays an irreplaceable role in autonomous driving and inteligent transportation systems. However，current segmentation networks often suffer from challenges such as blurred object boundaries，mutual occlusions between objects，and significant variations in object scales，which hinder segmentation accuracy. To address these issues，this paper proposed a city road scene semantic segmentation network that integrates global context awareness and multi-scale feature fusion.To mitigate the problem of blurred segmentation boundaries，a Global Awareness Module（GAM）was designed to enhance interaction between spatial and channel-wise information，enabling the network to capture comprehensive global context. For handling object occlusion and improving recognition of partially obscured regions，a Multi-Scale Feature Fusion Module（MSFF） was introduced，which efectively integrated contextual cues from different receptive fields to ensure segmentation accuracy for objects of varying sizes.Furthermore，a comprehensive Multi-constraint Feature Smoothing Loss was employed to enforce spatial coherence and semantic consistency，guiding the model toward a more optimal solution by refining feature distributions around object boundaries and within complex scenes.Extensive experiments are conducted on two benchmark datasets. On the Cityscapes dataset，the proposed method achieves mIoU improvements by 0.5% ， 0.9% ，and 1.7% under different input resolutions. On the ADE2OK dataset，an mIoU gain of （204号 2.1% is observed. Compared with existing semantic segmentation models，the proposed approach demonstrates superior performance in urban road scene understanding，particularly in terms of boundary delineation，and occlusion robustness.

Key words： deep learning；image processing; semantic segmentation； feature fusion； loss function

1引言

语义分割涉及对图像的像素级别分类以实现对多场景的精准理解，广泛应用于自动驾驶、医学图像分析、遥感影像分析等工程任务中[1]从开山之作全卷积神经网络 FCN[2] 后，Ronneberger等提出U-Net[3引入编解码结构和跳跃连接以改善上采样的空间信息保留，Google团队提出DeepLab[47系列模型，PSPNet8]聚合全局上下文信息，MaskR-CNN[9]在FasterR-CNN[10]的基础上加入并行分支以预测分割掩码，推进语义分割发展。（剩余18708字）