When Infrared Meets Text: Frequency-Enhanced Alignment with Uncertainty Quantification for Infrared Image-Text Retrieval

Infrared imaging extends human perception beyond the visible spectrum, yet it introduces an Infrared Semantic Contradiction (ISC), where physical thermal attributes fundamentally diverge from the descriptive semantics of natural language. This mismatch creates a critical bottleneck for cross-modal retrieval, as conventional models fail to extract discriminative features from infrared imagery due to its inherent semantic sparsity. To bridge this gap, we propose FIRE-UQ, a framework designed to robustly align text with noisy thermal cues. We incorporate a frequency-domain branch to recover structural details typically lost in the spatial domain, and introduce an uncertainty-aware contrastive learning mechanism to dynamically down-weight unreliable correlations. To further reinforce semantic discrimination, we integrate a saliency-guided negative mining strategy that compels the model to distinguish core thermal targets from environmental distractions. Furthermore, we establish the first comprehensive Infrared Image-Text Retrieval (II-TR) Benchmark via multi-source construction strategies. Extensive experiments demonstrate that FIRE-UQ significantly outperforms state-of-the-art baselines by effectively resolving the semantic challenges of the infrared domain.