联合注意力机制与多级特征融合的街景全景分割算法研究

Streetscape Panorama Segmentation Algorithm Combined with Attention Mechanism and Multi-level Feature Fusion

全景分割是场景认知领域的新型挑战性任务,针对此任务中的卷积神经网络和下采样的计算范式中,存在建模长距离特征依赖学习效率低和小物体信息丢失的局限性。本文设计了一种联合注意力机制和多级特征融合的街景分割算法。首先,使用压缩和激励网络(Squeeze-and-Excitation Networks,SENet)改进Panpotic-Deep-Lab网络,聚合通道上下文信息,加强网络学习重要特征,抑制无用特征。其次,在解码器阶段引入多尺度特征融合模块弥补部分物体信息。为验证方法的有效性,基于Cityscapes城市景观基准数据集,实验结果显示:本文方法提取的全景分割质量(PQ)为59.5%,分割质量(SQ)为80.5%,识别质量(RQ)为73.4%,平均交并比(mIOU)为78.8%。结果表明,本文模块可以有效提升全景分割精度。

Panoramic segmentation is a novel and challenging task in the field of scene perception,for which there are limitations of modelling long-range feature-dependent learning inefficiencies and small object information loss in the computational paradigm of con-volutional neural networks and downsampling.In this paper,a streetscape segmentation algorithm combined with attention mechanism and multi-level feature fusion is designed.Firstly,the Panpotic-DeepLab network is improved using Squeeze-and-Excitation Net-works(SENet)to aggregate channel context information,enhance the network to learn important features and suppress useless fea-tures.Secondly,a multi-scale feature fusion module is introduced in the decoder stage to compensate for some of the object informa-tion.To verify the effectiveness of the method,the experimental results based on the Cityscapes benchmark dataset show that:the panoramic segmentation quality(PQ)by this method is 59.5%,the segmentation quality(SQ)is 80.5%,the recognition quality(RQ)is 73.4%,and the average intersection and merge ratio(mIOU)is 78.8%.The results show that the module in this paper can effectively improve the panoramic segmentation accuracy.