Multiview Fine-Grained Contrastive Learning for Radar Open-Set Recognition
Abstract
Radar open-set recognition (OSR) requires models to effectively reject unknown classes while maintaining high precision in recognizing known targets. Although existing contrastive learning approaches can learn compact intraclass representations, the prevalent paradigm of aggregating features via global average pooling (GAP) often leads to information loss. We identify a critical flaw in this approach termed “structural amnesia,” which refers to the irreversible loss of high-frequency structural details after GAP (e.g., Micro-Doppler textures and scattering center distributions) that are essential for distinguishing fine-grained radar signals. To address this limitation, we propose multiview fine-grained contrastive learning (MVFCL). Unlike mainstream frameworks that compress signals into global vectors, MVFCL employs a dimension-aware mechanism to perform contrastive alignment on preserved high-dimensional tensors. Specifically, MVFCL incorporates two key components: 1) structure-preserving contrastive mechanism: a multiexpert backbone is employed to decouple signal features along physical dimensions. Driven by our dimension-aware dense contrastive loss, this mechanism performs fine-grained alignment on intact feature maps, preserving high-frequency details critical for anomaly detection and 2) dual-level mixed manifold regularization: we propose a shallow mixing strategy to counteract the semantic bias induced by radar spatial sparsity. This approach mixes features in the shallow space to mitigate label inconsistency caused by sparsity, ensuring that virtual samples maintain strict label fidelity to smooth the decision boundary. Extensive experiments demonstrate that MVFCL achieves state-of-the-art performance on measured high-resolution range profile, micro-Doppler spectrograms (waterfall graphs), and standard OSR benchmarks. Notably, it outperforms competitive baselines by margins of 1.3%–11.0% in area under the receiver operating characteristic (AUROC) across various datasets and achieves an AUROC of 89.2% on the challenging Tiny-ImageNet benchmark.