Infrared-Guided Reconstruction of Visible Images Occluded by Localized Smoke
In safety-critical scenarios such as fire rescue and autonomous driving, localized smoke severely degrades visible-light (VIS) image quality, making smoke removal and scene restoration a fundamental yet challenging task. Existing approaches suffer from modality conflicts when fusing infrared (IR) and visible features, failing to reconstruct authentic color images in smoke-occluded regions while preserving texture details in smoke-free areas. To address this, we propose SADB-Net, which leverages IR imaging to provide accurate structural priors for reconstructing VIS images across different smoke locations and densities. Specifically, we introduce a Cross-Modal Guidance Module (CMGM) that decouples IR structure extraction from VIS texture restoration to avoid modality conflicts. Within CMGM, a spatial adaptive gating mechanism selectively injects IR priors into occluded regions, complemented by a dual-edge structure guidance loss to enforce geometric consistency. In addition, we construct a benchmark dataset of 3,854 registered IR-VIS image pairs and evaluate against dehazing, inpainting, and fusion methods. Experiments show that SADB-Net yields visually coherent reconstructions, generalizes effectively to real-world conditions, and significantly enhances downstream object detection performance.