Bridging the Thermal Gap: Data-Model Collaborative Transfer for Infrared Perception
Infrared visual perception fundamentally differs from visible-domain vision due to its distinct thermal imaging mechanism and varying imaging conditions. Existing methods typically decouple data construction from model transfer, overlooking their interdependence: model accuracy and robustness depend critically on the coverage and composition of training observations, while effective model transfer requires explicit modeling of infrared-specific imaging characteristics. To address this issue, we propose a data-model collaborative transfer framework for infrared visual perception. At the data level, a task-responsive generator learns weights over a feasible infrared imaging-variation space to construct informative training observations. At the model level, a formation-guided representation module derives complementary radiance-structure and response-detail descriptors from infrared observations. We further design a thermal-aware adapter that uses these descriptors to modulate intermediate representations and produce lightweight residual corrections for model transfer. A single-loop training strategy coordinates data generation, representation learning, and adapter optimization without expensive nested optimization. Extensive experiments on image enhancement, depth estimation, small-target detection, and semantic segmentation demonstrate consistent performance improvements across multiple foundation-model backbones with a small trainable parameter budget.
Authors
- Benzhuang Zhang (ORCID: https://orcid.org/0009-0003-1755-8510)
- Zhirui Fang
- Jinyuan Liu (ORCID: https://orcid.org/0000-0003-2085-2676)
- Risheng Liu
- Zhu Liu
- Qilong Zhao
Institutions
- Dalian University of Technology (CN)
Publication Details
- Journal
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1109/tpami.2026.3754462
- Primary Topic
- Infrared Target Detection Methodologies
- Type
- article
- Field-Weighted Citation Impact
- 0.00