基于双分支融合的图像实时语义分割方法  

A Real-Time Image Semantic Segmentation Method Based on Dual Branch Fusion

作　　者：宋玉琴[1] 娄辉 张琪 商纯良 SONG Yuqin;LOU Hui;ZHANG Qi;SHANG Chunliang(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710600,China)

机构地区：[1]西安工程大学电子信息学院,西安710600

出　　处：《空军工程大学学报》2025年第2期62-70,共9页Journal of Air Force Engineering University

基　　金：中国纺织工业联合会科技指导性项目(2019062);陕西省教育厅专项科研计划项目(18JK0358)。

摘　　要：针对现有实时语义分割网络分割多尺度目标时存在类别错分和分割不完整的问题,提出了一种基于双分支融合的图像实时语义分割方法。提出尺度注意融合模块,融合细节分支和语义分支提取到的目标空间特征和语义信息,以提高网络对多尺度目标识别的准确率。使用边缘损失函数引导细节分支学习目标边缘轮廓,增强网络对目标边缘细节的分割性能。最后,构建全局感知模块提高网络的全局上下文感知能力。实验结果表明:文中方法在CityScapes和CamVid数据集上平均交并比(mIoU)分别为78.1%和76.2%,平均像素准确率(mPA)分别为87.6%和85.4%,对于小尺度目标边缘实现了更精准的分割,且在一个GTX 1080Ti GPU上推理达到实时要求,帧速率(FPS)分别达到59.8和43.5。Aimed at the problems that faulty classification and incomplete segmentation are in existence in segmenting multi-scale objects to the existing real-time semantic segmentation networks,a real-time semantic image segmentation method is proposed based on dual branch fusion.The method introduces a scale attention fusion module that is able to fuse object spatial feature and semantic information extracted from the detail branch and semantic branch,thereby improving the accuracy of the network for multi-scale object recognition.The edge loss function is used to guide the detail branch into learning the object edge contour,improving the network’s segmentation performance on object edge details.Finally,a global perception module is constructed to enhance the global context perception capability of the network.The experimental results demonstrate that the proposed method achieves the mean Intersection over union(mIoU)of 78.1%and 76.2%on the CityScapes and CamVid datasets respectively.Additionally,the mean pixel accuracy(mPA)is 87.6%and 85.4%,respectively.For small-scale object edges,there is a more accurate segmentation,coming up to the real-time requirements on a single GTX 1080Ti GPU,and frames per second(FPS)achieves 59.8 and 43.5 respectively.

关 键 词：深度学习 实时语义分割 尺度注意 特征融合 全局感知 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]