Hierarchical Sparse Reconstruction Based Multi-feature Saliency for Target Detection in SAR Images

The utilization of clutter information is useful for target detection. However, the clutter statistical modeling, used in conventional constant false alarm rate (CFAR) methods, is usually difficult for a real synthetic aperture radar (SAR) image. And though intensity feature is basic for target detection, only using intensity feature is insufficient considering the speckle noise. To address these problems, we propose a novel hierarchical sparse reconstruction based multi-feature saliency method to detect targets in SAR images. The proposed method hierarchically generates saliency maps using sparse reconstruction based on clutter templates on instance feature and structural feature separately. The clutter templates sampled from the pure clutter SAR images contains the clutter information. Though the sparse reconstruction based on clutter templates, we can fully utilize clutter information without clutter statistical modeling. Moreover, the proposed method jointly uses the instance feature and the structure feature, making the saliency detection more robust to speckle noise. We evaluate the proposed method on the miniSAR real data. The experimental results demonstrate that the proposed method outperforms conventional SAR target detection methods in terms of the precision, recall and F1-score.