FSAR-Cap：Large-scale fine-grained SAR image captioning dataset

The FSAR Cap dataset aims to build an image text reference corpus with fine semantic description capabilities for SAR image semantic understanding and cross modal modeling, promoting the development of SAR image automatic interpretation, image captioning, and remote sensing multimodal models. This dataset is constructed based on the FAIR-CSAR object detection dataset, consisting of 14480 SAR images and 72400 accompanying descriptive texts. FSAR Cap adopts a two-stage annotation method: first, it utilizes detection results and spatial location information to automatically generate basic descriptions through multiple templates; Then, combined with manual verification and language model polishing. In the end, each image will generate 5 descriptive sentences with different styles and complementary information, covering target types, quantities, positional relationships, external features, etc. As the first large-scale semantic annotation dataset with fine-grained description hierarchy for SAR images, FSAR Cap not only improves the semantic expression quality of SAR images, but also provides a unified and high-quality data benchmark for image captioning, remote sensing visual language model training, multimodal inference, and SAR natural language alignment research in the SAR field, laying the foundation for the further development of SAR automated interpretation and intelligent semantic understanding technology system.