计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (20): 188-196.DOI: 10.3778/j.issn.1002-8331.2203-0122

• 图形图像处理 • 上一篇    下一篇

基于边缘增强的遥感图像弱监督语义分割方法

栾晓梅,刘恩海,武鹏飞,张军   

  1. 1.河北工业大学 人工智能与数据科学学院,天津 300401
    2.北京仿真中心 航天系统仿真重点实验室,北京 100854
  • 出版日期:2022-10-15 发布日期:2022-10-15

Weakly-Supervised Semantic Segmentation Method of Remote Sensing Images Based on Edge Enhancement

LUAN Xiaomei, LIU Enhai, WU Pengfei, ZHANG Jun   

  1. 1.School of Artificial Intelligence, Hebei University of Technology, Tianjin 300401, China
    2.Key Laboratory of Aerospace System Simulation, Beijing Simulation Center, Beijing 100854, China
  • Online:2022-10-15 Published:2022-10-15

摘要: 随着弱监督学习被应用于遥感图像语义分割,大大降低了模型训练的数据成本。然而,由于监督信息不足,类激活图难以准确激活出遥感图像中不同尺度大小的目标,这使得基于类激活图获得的伪分割掩码边缘粗糙,从而导致最终的分割结果不准确。此外,大部分的弱监督语义分割方法都是基于可视化的两阶段方法,模型复杂繁琐。针对上述问题,设计了一种基于边缘增强的端到端弱监督语义分割网络。在特征空间边缘增强模块中,以自监督方式引导网络学习遥感图像中尺寸不一的目标,并且,细化伪分割掩码的边缘;在输出空间边缘增强模块中,通过端到端训练提升分割精度,同时降低模型训练的繁琐度。在ISPRS 2D数据集上的实验结果表明,该方法在仅使用图像级标签的情况下MIoU分别为57.72%和59.45%,与其他方法相比,效果较好。

关键词: 弱监督学习, 遥感图像语义分割, 自监督学习, 边缘增强, 端到端

Abstract: With the application of weakly supervised learning in remote sensing image semantic segmentation, the data cost of model training is greatly reduced. However, due to the lack of supervision information, the class activation map is difficult to accurately activate the targets with different scales in the remote sensing image, which makes the edge of the pseudo segmentation mask obtained based on the class activation map rough, resulting in the inaccurate final segmentation result. In addition, most of the weakly supervised semantic segmentation methods are two-stage methods based on visualization, and the model is complex and cumbersome. To solve the above problems, this paper designs an end-to-end weakly supervised semantic segmentation network based on edge enhancement. Firstly, this paper proposes a feature space edge enhancement module, which guides the network to learn the targets with different sizes in the remote sensing image in a self-monitoring way, and refines the edge of the pseudo segmentation mask. Secondly, this paper designs the output space edge enhancement module. Through end-to-end training, it not only improves the segmentation accuracy, but also reduces the complexity of model training. The experimental results on the ISPRS 2D dataset show that the MIoU of this method is 57.72% and 59.45% respectively when only image level labels are used, which is better than other methods.

Key words: weakly supervised learning, semantic segmentation of remote sensing image, self-supervised learning, edge enhancement, end-to-end