- https://doi.org/10.1109/icsidp62679.2024.10868342
Few-Shot Object Detection in SAR images via Gaussian Flow Class-Balanced Differential Aggregation
- Nov 22, 2024
- Po Zhao +3 more
Currently, few-shot object detection (FSOD) methods for synthetic aperture radar (SAR) images are severely understudied. The high confidentiality and scarcity of SAR data, coupled with the susceptibility of targets in SAR images to significant changes in orientation, make existing models sensitive to variations in novel class examples. This sensitivity leads to challenges in extracting class prototypes and accurately estimating class centers for novel classes. To solve the above issues, we propose a few-shot object detection algorithm for SAR images via Gaussian Flow Class-Balanced Differential Aggregation. Specifically, we first design a class-balanced difference feature aggregation module to solve the problems of difficult class distribution estimation and inaccurate class center estimation of novel classes in SAR images. Second, the contextual information perception module is designed to extract the most representative support features into fine-grained prototypes and aggregate them with the query features using the matching paradigm to solve the imbalance between foreground targets and background in SAR images. Finally, experimental results on MSAR-AIR dataset indicate that our model realizes the best detection properties on SAR images compared to existing models.
- # Synthetic Aperture Radar Images
- # Detection In Synthetic Aperture Radar Images
- # Targets In Synthetic Aperture Radar Images
- # Synthetic Aperture Radar
- # Foreground Targets
- # High Confidentiality
- # Query Features
- # Changes In Orientation
- # Object Detection In Images
- # Object Detection Methods