T-SFANet: A spatial-frequency attention network for thermal infrared image dehazing

Abstract

Thermal infrared (TIR) image dehazing aims to mitigate degradation caused by haze-induced scattering and absorption, thereby restoring structural clarity and detail visibility. Previous dehazing approaches have been developed for visible images, and their direct application to TIR data is limited due to fundamental differences in radiation characteristics and texture distribution. To this end, T-SFANet is proposed as a spatial-frequency attention network specifically designed for TIR image dehazing. A Local-Global Alternating Attention block is designed, which alternates between local window attention and sparse global attention to simultaneously preserve fine textures and capture large-scale haze structures. In addition, a Frequency-domain Self-Attention mechanism and a Band FFT Loss are incorporated to enhance the recovery of mid- and high-frequency components, improving detail and edge reconstruction. Furthermore, a real-world dataset named TIR-HAZE is constructed, representing the first TIR dehazing benchmark collected under natural haze conditions. Experiments on both M3FD-Synthetic and TIR-HAZE demonstrate that T-SFANet achieves superior restoration accuracy compared to state-of-the-art methods in terms of visual quality and quantitative evaluation.