Distribution-Aware Infrared Small Target Detection Based on Multiscale Convolutional Decoder and Hypergraph Attention

Infrared small target detection (IRSTD) has advanced substantially with the rapid progress of deep learning, yet prevailing approaches—predominantly based on local convolution or attention-driven feature enhancement—remain inadequate for modeling high-order correlations among multiscale features, thereby leading to persistent false alarms and missed detections under complex backgrounds and low signal-to noise ratio (SNR) conditions. To overcome these limitations, we present an IRSTD network that couples a multiscale convolutional decoder (MSCD) with a hypergraph attention (HA) mechanism and introduces a distribution-aware pixel-wise Wasserstein loss (DPWL) to jointly enforce scale, positional, and distributional consistency. Concretely, MSCD aggregates cross-layer representations through multireceptive-field convolutional structures to preserve fine-grained target details and amplify saliency; HA employs learnable soft hyperedges to capture high-order dependencies across channel and spatial dimensions, suppressing background interference via global contextual reasoning; and DPWL injects a distribution-consistency constraint into the optimization process to enhance convergence robustness and localization accuracy in SNR condition scenarios. Extensive experiments on four datasets (IRSTD-1k, NUDT-SIRST, DenseSIRST, and NUAA-SIRST) demonstrate consistent gains over state-of-the-art methods across all metrics, namely intersection over union (IoU), P d , and F a . Specifically, the proposed model attains an impressive average performance of 75.89% in IoU, 96.70% in P d , and 8.67× 10 −6 in F a across these evaluation benchmarks. Finally, a comprehensive ablation study corroborates the effectiveness and complementary roles of the three core components.