Sonar Image Coarse-to-Fine Few-Shot Segmentation Based on Object-Shadow Feature Pair Localization and Level Set Method

Sonar image segmentation is the key step in sonar image analysis automatically. Due to the scarcity of sonar samples and the uneven seabed background, it is very difficult to distinguish underwater object’s contour accurately. In this article, a novel two-stage coarse-to-fine object segmentation model is proposed. In the first stage, the proposed model utilizes object-shadow features to locate objects in sonar images, and in the second stage, the level set algorithm evolves from a coarse curve to a fine object contour. To address the few-shot issue of sonar samples, by analyzing the common features between sonar, optical, and remote sensing images, we migrate samples with object-shadow pair features from optical and remote sensing fields to augment sonar image datasets. The ablation study has demonstrated the effectiveness of such migration between heterogeneous data. In the bounding box selection step of object detection, an algorithm is designed based on the causal association between object and shadow, so that the model can remove the isolated region and output the unique object-shadow pair. Extensive experimental results prove the proposed algorithm is effective and robust by comparing it with five other segmentation algorithms based on the level set method.