| Title |
Spatial Adaptive Kernel Network for Enhanced Oriented Ship Detection in Synthetic Aperture Radar Imagery |
| DOI |
https://doi.org/10.5573/ieie.2024.61.9.78 |
| Keywords |
Oriented object detection; Synthetic aperture radar; SAR ship detection |
| Abstract |
The existing oriented object detection models have shown excellent performance with optical images but have demonstrated inferior detection performance with Synthetic Aperture Radar (SAR) ship images. This is particularly evident in inshore areas with complex backgrounds, where the models exhibit vulnerabilities. To address this issue, we propose the Spatial Adaptive Kernel Network (SAKNet), a ship detection system specifically designed for SAR images. SAKNet is tailored for oriented ship detection in SAR imagery. We observed that the geographical characteristics of inshore areas are significantly different from coastal regions, which led us to develop a Spatial Adaptive Kernel based on these unique features to overcome detection weaknesses in these areas. The spatial adaptive kernel divides SAR images into small grid-like regions and adaptively applies the local statistics of these subdivided grid areas to the receptive field. This approach enhances the spatial information utilization capability of the detection kernel, thereby improving ship detection performance in coastal regions. SAKNet demonstrated superior performance on both the Official-SSDD and HRSID datasets, with a notable increase in detection accuracy in inshore areas. Extensive experiments have proven that SAKNet is effective for oriented ship detection (OSD) in SAR images. |