Less Labeling Burden: A SAR Ship Detection Data Selection Method Based on K-means

In deep learning-based object detection, labeled data is crucial. However, annotating ship targets in Synthetic Aperture Radar (SAR) images is challenging because of the distinct characteristics of SAR images, especially when employing Oriented Bounding Boxes (OBB). These challenges arise from factors such as varying ship orientations, complex backgrounds, and the presence of noise, making accurate annotation more difficult. To reduce the labeling burden while minimizing its impact on ship detection performance, we propose a SAR data selection method. This method ensures that, despite limited quantities, the data contains a more representative range of scenarios when constructing a new dataset. By selecting 10 features of SAR images and performing K-means clustering based on these features, we choose the samples closest to the cluster centers as the selected data. We have validated the effectiveness of the proposed method by applying two typical object detection models to both randomly selected and clustering-selected datasets, thereby demonstrating the efficacy of proposed approach.