Unsupervised Ship Detection in SAR Images Based on Fusion of Statistical and Structural Features
Hanrui Shi, Shuqian Zhou, Tao Yang, Zuowei Zhang · 2024
SAR Ship detection plays an important role in ocean monitoring. For deep learning-based detection methods, it often requires a large number of accurate target-level annotations, which are usually hard to obtain. In this paper, an unsupervised ship target detection method based on fusion of statistical and structural features is proposed to deal with unlabeled synthetic aperture radar (SAR) image data. An iterative optimization detection strategy is developed to improve the accuracy of the parameter estimation of the gamma distribution statistical model in the constant false alarm rate (CFAR) detector. First, a large number of bounding boxes with false alarms are obtained with CFAR detector. Then, the knowledge distillation model is designed to increase the inter class gap between the clutter and target in the bounding boxes. The soft label information provided by the CFAR detector is employed to constrain the training of the Faster-RCNN detector. The bounding boxes generated by FasterRCNN serve as adaptive detection units within CFAR detector. After several iterations, the CFAR detector and the deep network detector are mutually optimized. The iteration is stopped when the total detection rate of the target and clutter falls below a preset threshold. Finally, the optimized Faster-RCNN detector model is used for target detection of other SAR image data. The effectiveness of the proposed method on SAR ship detection is demonstrated by some experiments with real SAR images datasets without annotations.