Optimizing latent space for effective radar target detection using variational auto-encoder

Abstract

Detecting small floating marine targets is a significant challenge in radar systems, as conventional neural networks fail to detect targets effectively due to the lack of discriminative prior information. To address this issue, this paper proposes a prior-guided, weakly supervised detector based on a multi-scale temporal variational auto-encoder (MST-VAE). First, radar returns are represented as one-dimensional sliding window Doppler sequences (SWDS) to enhance clutter–target separability. Then, an encoder with multi-scale and dilated convolutions is designed to match the Doppler irregularity of sea clutter and the periodic Doppler spikes of target returns in the SWDS. In addition, two clutter-focused loss functions are developed to ensure the model focuses on learning clutter properties without overfitting to simulated targets. Finally, three complementary anomaly scores are extracted from the MST-VAE and fused in a fast convex-hull detector. Experiments on measured radar data demonstrate that the proposed method outperforms a strong feature-based baseline, with average and maximum detection performance gains of 5.2% and 16.9%, respectively.