Answers are grounded in this paper and cite their sources.
Infrared small target detection (IRSTD) has advanced substantially in recent years. Nevertheless, infrared images contain limited discriminative information, as targets exhibit weak texture cues and are easily corrupted by complex backgrounds. As a result, methods that rely solely on visual features are gradually approaching a performance bottleneck. To address this issue, we introduce both language priors and structural priors into IRSTD and propose a dual-prior-guided multiscale network (DPMSNet). Specifically, we design a structural saliency prior module to explicitly characterize local contrast and spatial structure using fixed-parameter operators, thereby enhancing target saliency and suppressing background noise. We further introduce a language prior modulation module, which provides global semantic guidance through structured scene descriptions and requires no textual input during inference. In addition, a progressive multiscale feature extraction module is developed to capture target responses at different scales, and an adaptive scale fusion strategy is adopted to effectively integrate multiscale information. Extensive experiments show that DPMSNet achieves strong overall performance on the NUAA-SIRST, NUDT-SIRST, and IRSTD-1 K datasets, with probability of detection ( Pd ) values of 97.710%, 99.259%, and 94.613%, respectively.
No takes yet. Share an insight, caveat, or question.
Tianci et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: