Domain Adaptive SAR Ship Detection based on Knowledge Distillation and Transformer Attention

Synthetic Aperture Radar (SAR) ship detection has garnered considerable interest owing to its versatile application in both civilian and military field. Current SAR ship detectors based on Convolutional Neural Networks (CNNs) exhibit remarkable detection performance, contingent upon the availability of sufficient labeled SAR samples. However, it is time consuming and artificially expensive to label the training SAR images. Moreover, when the conditions of training data such as sensors, platforms, imaging parameters change, it is inevitable to alter the feature distribution between the source domain and the target domain, leading to the deterioration in detection performance. Consequently, the development of a detector capable of learning domain-invariant features without large-scale labeled samples in target domain remains a significant challenge. In this paper, we construct a semi-supervised domain-adaptive SAR ship detector based on YOLOv5, incorporating the Mean Teacher model with a knowledge distillation framework to capture domain-invariant features and mitigate domain discrepancies. Additionally, we integrate the transformer encoder module and convolutional block attention module into prediction head to enhance the capability of the detector to identify SAR ships of various scales across different domains. Extensive experiments conducted on several SAR ship datasets obtained under diverse conditions validate the effectiveness of our proposed approach.