Cascaded Fine Detection of Small Object in Large-Scale SAR Images Leveraging AIS Data

Detecting small objects in large-scale Synthetic Aperture Radar (SAR) images presents significant challenges due to their subtle features and the interference from complex backgrounds. This is particularly problematic when identifying maritime vessels, as SAR imagery often includes extensive geographic information, with some land features resembling vessels. These similarities increase the likelihood of false detections, complicating the accurate identification and tracking of vessels. To address these challenges, this paper proposes a cascade fine detection methodology for small objects in large-scale SAR images, leveraging Automatic Identification System (AIS) data for enhanced accuracy. Specifically, to address the issue of small object detection in large-scale scenes, we design a remote sensing detection network (RSDNet), which optimizes feature extraction and multi-scale information fusion to capture more refined object details. To minimize false detections, AIS data is embedded into the detection framework as prior knowledge through cross-modal data fusion. Thus, AIS data is integrated with SAR images in both spatial and temporal dimensions. Meanwhile, we validate RSDNet's accuracy using three publicly accessible datasets as well as a self-built dataset from a remote sensing object detection vessel (RSOD vessel). Furthermore, to assess the efficacy of our approach, we create a SAR-AIS matching dataset in the absence of a benchmark dataset for AIS and SAR matching. Finally, the extensive experimental results indicate that the proposed methodology can provide promising detection of small objects in large-scale SAR images.