A unified progressive fusion network with adaptive degradation prompts for multi-weather thermal infrared image restoration

Thermal infrared (TIR) image restoration aims to recover clearer and more informative images from inputs degraded by unfavorable weather conditions. Current TIR restoration approaches, however, mainly focus on individual degradation categories, such as rain, snow, or haze, which limits their ability to address different weather degradation categories within a unified framework. To address this challenge, we introduce a TIR unified prompt-guided fusion network, an adaptive degradation-prompting framework designed for unified TIR image restoration under individually occurring rain, snow, or haze degradations. A weather-adaptive dynamic prompt module is first constructed, where a lightweight degradation predictor estimates the probability distribution over predefined degradation categories in TIR images and produces dynamic degradation prompts from a learnable prompt bank. The generated degradation prompts are then provided to the progressive fusion enhancement module and integrated through a prompt fusion strategy to guide feature enhancement and restoration, allowing the network to restore TIR images effectively under diverse weather conditions. To improve restoration quality across the evaluated degradation categories, we further design a joint loss function that combines Charbonnier reconstruction loss, Laplacian edge loss, and degradation-aware auxiliary constraints. Experimental results on the M3FD-based synthetic benchmark dataset demonstrate that the proposed framework achieves the best overall performance and competitive or superior performance across most evaluated degradation conditions, while effectively suppressing degradations caused by rain, snow, and haze. Moreover, the restored images achieve competitive performance in downstream object detection, further demonstrating the potential of the proposed framework for adverse-weather TIR perception. The source code is available at https://github.com/xia11998/TIR-UPF .