HRRSODN: high-Resolution remote sensing object detection network fusion of road-adjacent relationships

• Published in: Geocarto international Vol. 38; no. 1
• Main Authors: Xia, Liegang, Su, Yishao, Liu, Ruiyan, Mi, Shulin, Yang, Dezhi
• Format: Journal Article
• Language: English
• Published: Taylor & Francis 31.12.2023; Taylor & Francis Group
• Subjects: attention-mechanism; road distance loss; Road-adjacent; spatial relationships
• ISSN: 1010-6049; 1752-0762
• DOI: 10.1080/10106049.2023.2280549

Application of deep learning technology in remote sensing image target detection(RSITD)has made remarkable progress in recent years. However, existing methods usually fuse spatial context information at the pixel level while ignoring the mining of spatial relationships between ground objects. This paper proposes a high-resolution remote sensing object detection network fusion of road-adjacent relationships (HRRSODN) that integrates road proximity relations. Taking artificial features such as toll stations adjacent to roads as an example, the spatial relationship of road proximity is used to assist visual information and improve the extraction accuracy. The method optimizes detection by: Using the spatial relationship as prior information; fusing coordinate attention that captures the spatial relationship between objects; and adding the attention generated by the spatial relationship to the attention generated in the original X and Y directions. This enables the model to understand the spatial layout of objects in the image. The distance loss function concretizes the spatial relationship into a distance index. The distance loss is incorporated into the training process by measuring the spatial distance between feature targets. Thus, during training, the network focuses not only on the positional information of the targets but also pays more attention to the spatial layout of the features. Experimental results show that, compared with conventional methods, the proposed method has achieved significant improvements in detection performance and mining spatial relationships. This verifies the importance of considering the spatial relationship of ground objects in remote sensing target detection. Further, this paper proposes a new method for target detection in remote sensing images (RSIs), which effectively utilizes the spatial relationship information between targets, and provides a new idea for further improving target detection in RSI.