Multiperspective Visual Feature Enhancement and Interaction for Remote Sensing Image Captioning
Year: 2026
Type: article
Abstract: Recently, remote sensing image captioning (RSIC) has gained increasing attention in remote sensing images interpretation tasks. Remote sensing images (RSIs) exhibit significant scale variations, high intra-class variance, and low inter-class variance, which together pose substantial challenges to learn more discriminative visual representations and align visual features with text features. Although the fusion of grid-level and region-level features provide a relatively feasible solution, it's still valuable to explore more effective strategy. In this paper, we propose a novel RSIC method that aims to enhances the visual representation of RSIs while simultaneously improving the interaction between visual and textual representations. The proposed framework consists of two key components: Dual Branch Feature Enhancement (DBFE) module and Mask-Topic (MT) module. The DBFE module effectively integrates global contextual features of RSIs with locally salient features that exhibit strong scene relevance, thereby enriching the visual representations for subsequent decoding stages. Meanwhile, the MT module integrates scene-level semantic guidance into the text decoding process to enhance the model's robustness in handling incomplete or partially missing inputs. Finally, three widely used remote sensing captioning benchmarks are employed for performance evaluation. The experimental results demonstrate the effectiveness of the proposed method, which outperforms state-of-the-art approaches with superior performance. Source code: https://github.com/echoecho1023/Multi perspective-Visual-Feature-Enhancement
Institution: Xidian University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7155561117