Abstract:Visual grounding aims to locate target objects in images based on natural language descriptions, which is crucial for promoting the development of human-computer interaction technology. In recent years, visual grounding tasks in remote sensing images have gradually caught attention. Compared with visual grounding in natural images, remote sensing visual grounding faces numerous new challenges, such as more complex textual descriptions, a smaller proportion occupied by target objects in the images, and greater reliance on global contextual information in localization processes. To this end, this study proposes a remote sensing visual grounding method based on progressive fusion multimodal and multi-scale. By dynamically adjusting the weights of textual features during multimodal fusion, the proposed method gradually fuses contextual information to enhance the ability to comprehend complex textual descriptions. Meanwhile, deep features are utilized to guide the multi-scale fusion of shallow features, thereby supplementing detailed information to improve localization precision on the basis of fully preserving multimodal semantic information. Experiments are conducted on two public datasets, RSVG and DIOR-RSVG, and the experimental results demonstrate that the proposed method has significant advantages over existing methods.