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 ›› 2021, Vol. 47 ›› Issue (9): 203-209,216. doi: 10.19678/j.issn.1000-3428.0058447

• 图形图像处理 • 上一篇    下一篇

基于注意力机制和辅助任务的语义分割算法

叶剑锋, 徐轲, 熊峻峰, 王化明   

  1. 南京航空航天大学 机电工程学院, 南京 210008
  • 收稿日期:2020-05-27 修回日期:2020-08-27 发布日期:2020-09-10
  • 作者简介:叶剑锋(1984-),男,副教授、博士研究生,主研方向为图像识别;徐轲,本科生;熊峻峰,硕士研究生;王化明(通信作者),教授、博士生导师。
  • 基金资助:
    国家自然科学基金(61363066)。

Semantic Segmentation Algorithm Based on Attention Mechanism and Auxiliary Task

YE Jianfeng, XU Ke, XIONG Junfeng, WANG Huaming   

  1. College of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210008, China
  • Received:2020-05-27 Revised:2020-08-27 Published:2020-09-10

摘要: 为提高网络模型低层特征的离散度和语义分割算法的性能,以全卷积神经网络作为基础模型,提出一种基于辅助损失、边缘检测辅助任务和注意力机制的语义分割算法。通过重新设计网络模型的辅助损失分支,使网络低层特征编码更多语义信息。在多任务学习中,选择边缘检测作为辅助任务,基于注意力机制设计边缘检测的辅助任务分支,使网络模型更关注物体的形状和边缘信息。在此基础上,将基础模型、辅助损失分支、辅助任务分支集成构造为语义分割模型。在VOC2012数据集上的实验结果表明,该算法的平均交并比为71.5%,相比基础模型算法提高了6个百分点。

关键词: 注意力机制, 辅助任务, 辅助损失, 多任务学习, 语义分割

Abstract: When applied to semantic segmentation, the existing convolutional neural network models suffer from the low dispersion of low-level features, and thus reduce the performance of semantic segmentation algorithms.To address the problem, a basic fully convolutional neural network model is redesigned.On this basis, a novel semantic segmentation algorithm based on auxiliary loss, auxiliary edge detection tasks and attention mechanism is proposed.The auxiliary loss branch of the neural network model is redesigned to allow the low-level features to encode more semantic information.Then in multi-task learning, edge detection is chosen as the auxiliary task.The auxiliary task branch is designed based on the attention mechanism for edge detection to allow the network model pay more attention to the shape and edge information of objects.Finally, the basic model, auxiliary loss branch and auxiliary task branch are integrated into the semantic segmentation model.The experimental results on the VOC2012 dataset show that the proposed algorithm improves the mean intersection-over-union to 71.5%, outperforming the basic model algorithm by 6 percentage point.

Key words: attention mechanism, auxiliary task, auxiliary loss, multi-task learning, semantic segmentation

中图分类号: