CS-CapsFPN: A Context-Augmentation and Self-Attention Capsule Feature Pyramid Network for Road Network Extraction from Remote Sensing Imagery
CS-CapsFPN: A Context-Augmentation and Self-Attention Capsule Feature Pyramid Network for Road Network Extraction from Remote Sensing Imagery
复制标题
DOI：
10.1080/07038992.2021.1929884
复制
发表时间：
2021-05
影响因子：
2.6
通讯作者：
Yongtao Yu;Jun Wang;H. Guan;Shenghua Jin;Yongjun Zhang;Changhui Yu;E. Tang;Shaozhang Xiao
中科院分区：
文献类型：
--
作者：
Yongtao Yu;Jun Wang;H. Guan;Shenghua Jin;Yongjun Zhang;Changhui Yu;E. Tang;Shaozhang Xiao
摘要信息准确的道路网络数据库是非常重要的，并提供了许多交通相关活动的基本输入。近年来，遥感影像已成为道路网快速更新的重要数据源。然而，由于遥感图像中道路存在遮挡、阴影、材质变化、拓扑变化等多种复杂场景，实现高精度的道路提取仍然是一个难点。提出了一种新的上下文增强自注意胶囊特征金字塔网络（CS-CapsFPN），用于遥感图像中道路的提取。通过设计一种胶囊特征金字塔网络结构，CS-CapsFPN可以提取和融合不同层次和不同尺度的高阶胶囊特征，为道路区域地图预测提供高分辨率和强语义的特征表示。通过集成上下文增强和自我注意模块，所提出的CS-CapsFPN可以在高分辨率的角度利用多尺度上下文属性，并强调通道信息特征，以进一步增强特征表示的鲁棒性。在两个测试数据集上的定量评价表明，所提出的CS-CapsFPN实现了具有竞争力的性能，精确度，召回率，交集超过联合，和F得分分别为0.9470，0.9407，0.8957，和0.9438。比较研究也证实了所提出的CS-CapsFPN在道路提取任务的可行性和优越性。
Abstract The information-accurate road network database is greatly significant and provides essential input to many transportation-related activities. Recently, remote sensing images have been an important data source for assisting rapid road network updating tasks. However, due to the diverse challenging scenarios of roads in remote sensing images, such as occlusions, shadows, material diversities, and topology variations, it is still difficult to realize highly accurate extraction of roads. This paper proposes a novel context-augmentation and self-attention capsule feature pyramid network (CS-CapsFPN) to extract roads from remote sensing images. By designing a capsule feature pyramid network architecture, the proposed CS-CapsFPN can extract and fuze different-level and different-scale high-order capsule features to provide a high-resolution and semantically strong feature representation for predicting the road region maps. By integrating the context-augmentation and self-attention modules, the proposed CS-CapsFPN can exploit multi-scale contextual properties at a high-resolution perspective and emphasize channel-wise informative features to further enhance the feature representation robustness. Quantitative evaluations on two test datasets show that the proposed CS-CapsFPN achieves a competitive performance with a precision, recall, intersection-over-union, and F score of 0.9470, 0.9407, 0.8957, and 0.9438, respectively. Comparative studies also confirm the feasibility and superiority of the proposed CS-CapsFPN in road extraction tasks.