作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (4): 247-252,259. doi: 10.19678/j.issn.1000-3428.0054245

• 图形图像处理 • 上一篇    下一篇

基于密集层和注意力机制的快速语义分割

程晓悦, 赵龙章, 胡穹, 史家鹏   

  1. 南京工业大学 电气工程与控制科学学院, 南京 211816
  • 收稿日期:2019-03-15 修回日期:2019-06-03 出版日期:2020-04-15 发布日期:2019-07-12
  • 作者简介:程晓悦(1995-),女,硕士研究生,主研方向为计算机视觉、图像处理;赵龙章,教授、博士;胡穹、史家鹏,硕士研究生。
  • 基金资助:
    国家自然科学基金(61403189)。

Fast Semantic Segmentation Based on Dense Layer and Attention Mechanism

CHENG Xiaoyue, ZHAO Longzhang, HU Qiong, SHI Jiapeng   

  1. College of Electrical Engineering and Control Science, Nanjing Tech University, Nanjing 211816, China
  • Received:2019-03-15 Revised:2019-06-03 Online:2020-04-15 Published:2019-07-12

摘要: 针对传统语义分割网络速度慢、精度低的问题,提出一种基于密集层和注意力机制的快速场景语义分割方法。在ResNet网络中加入密集层和注意力模块,密集层部分采用两路传播方式,以更好地获得多尺度目标,并使用分组卷积减少计算量。同时在特征提取网络中加入注意力模块,以减少精度损失。实验结果表明,该方法在保证分割精度的前提下提升了分割速度,在Cityscapes数据集上得到了81.5%的MIOU,速度为42.3 frame/s,在ADE20K数据集上得到了61.8%的MIOU,速度为27.9 frame/s。

关键词: 语义分割, 轻量级网络, 分组卷积, 密集层, 注意力机制

Abstract: To address the low speed and accuracy of existing semantic segmentation networks,this paper proposes a fast scenario semantic segmentation method based on the dense layer and attention mechanism.The method adds the dense layer and attention module into ResNet.The dense layer adopts two-channel transmission that helps in obtaining multi-scale targets.It also uses group convolution to reduce the amount of computation.An attention module is introduced into the feature extraction network to reduce the loss of accuracy.Experimental results show that the proposed method gets an MIOU of 81.5% and a speed of 42.3 frame/s on the cityscape dataset.It also gets an MIOU of 61.8% and a speed of 27.9 frame/s on the ADE20K dataset,which means the proposed method can improve the segmentation speed while keeping the accuracy.

Key words: semantic segmentation, lightweight network, group convolution, dense layer, attention mechanism

中图分类号: