基于多分支结构和注意力机制的实时语义分割网络  

Real-time semantic segmentation network based on multi-branch structure and attention mechanism

作　　者：曾永煌 张孙杰 ZENG Yonghuang;ZHANG Sunjie(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《智能计算机与应用》2024年第5期107-114,共8页Intelligent Computer and Applications

基　　金：国家自然科学基金(61673276,61603255)。

摘　　要：在实时语义分割方法研究中,由于目标感受野有限,目前仍然存在大目标分割不准确和细节信息丢失的问题。针对这个问题,提出一种基于多分支结构和注意力机制的实时语义分割算法。首先,本文构建多分支结构的细节路径以保留多尺度细节信息,减少小目标细节丢失;其次,设计空洞金字塔分支扩大感受野,以覆盖视野内大目标,进一步丰富上下文信息;最后,提出双边注意力特征融合模块,以增强特征融合时对关键特征的选择,弥补小目标信息的缺失。在Cityscapes测试集、CamVid测试集所提模型的平均交并比(mIoU)为74.6%与73.6%,每秒传输帧数(Frames Per Second,FPS)为94与74;较于BiSeNet,本文算法的mIoU分别提高了6.2、8.0个百分点。实验结果表明,本文算法在实时性和准确性方面获得了很好的平衡。To address the problem of limited receptive fields in current real-time semantic segmentation methods leading to inaccurate segmentation of large objects and loss of detail information,this paper proposes a real-time semantic segmentation algorithm based on multi-branch structure and attention mechanism.First of all,design detail path of multiple branch structures to preserve multi-scale detail information and reduces the loss of small target details;Secondly,design the atrous pyramid branch to expand the receptive field and cover large targets within the field-of-view,thereby enriching context information;Finally,design a bilateral attention feature fusion module to enhance the selection of key channels during feature fusion and compensate for the missing of small target information.Experimental results on Cityscapes test set and CamVid test set show that the mean Intersection over Union(mIoU)of the proposed model is 74.6%and 73.6%,Frames Per Second(FPS)is 94 and 74.In comparison with BiSeNet,mIoU of the proposed model is increased by 6.2 and 8.0 percentage points respectively.Experimental results show that the algorithm proposed in this paper has achieved a good balance between real-time performance and accuracy.

关 键 词：实时语义分割 多分支结构 注意力机制 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]