Cross-Domain Few-Shot SAR Ship Object Detection with Meta-RCNN

In recent years, the advent of deep learning technology has significantly enhanced the efficacy of ship target detection and recognition in Synthetic Aperture Radar (SAR) imagery. However, in certain specific scenario, SAR ship images are extremely difficult to obtain. Moreover, the data distribution of these samples markedly differs from that of other SAR images, rendering the direct use of other large-scale SAR datasets for model training and generalization impractical. Most Few-shot object detection methods aim to locate and classify targets of novel classes in data-scarce scenerios, but their performance diminishes when confronted with data from unseen domains. To achieve few-shot object detection in images with novel distributions, rather than just novel classes, we meticulously constructed the CrossDomainShipDataset (CDSD), which comprises optical and SAR ship images from widely-used open-source datasets. A limited number of SAR ship images with specific distributions, obtained through our simulations, were used as benchmarks for evaluation. We utilized Meta-RCNN as the base model and integrated a pluggable Class-Domain Prototypes Extraction (CDPE) module to generate prototypes of class and domain at the same time. The model was trained on images from base domains to develop the capacity of learning to learn various distributions during the base training stage, then was fine-tuned with 1, 3, 5, and 10-shot configurations. Benchmark experiments demonstrate the efficacy of our approach for cross-domain few-shot SAR ship object detection.