Segmentation-Based Radar Target Detection With Cell-Wise Contrastive U-Net
Abstract
Radar target detection in strong clutter environments remains a challenging problem. This article proposes a segmentation-based radar target detection framework to improve the discrimination between targets and clutter. Specifically, the radar detection task is reformulated as a binary segmentation problem in the range-pulse domain, enabling the neural network to effectively capture the structural characteristics of radar returns. The proposed framework is implemented on a U-Net backbone to exploit multiscale feature representations for detection. To further enhance segmentation capability and construct a more discriminative feature space, a cell-wise contrastive learning strategy is introduced by integrating contrastive loss with the segmentation loss during training. In addition, a global context-aware classification module (GCACM) is designed to rapidly identify pure clutter samples, thereby reducing the computational cost during detection. The experiments conducted on both simulated and measured data from the SDRDSP database demonstrate that the proposed method achieves superior detection performance compared with the existing approaches.