CTSC: Infrared image colorization with topology-aware semantic consistency

Unsupervised infrared image colorization aims to convert unpaired single-channel infrared images into multi-channel color images that align with human visual perception. While existing methods use patch-wise contrastive learning to explore semantic correspondences, they overlook implicit topology-aware semantic relationships, resulting in texture distortion and detail blurring. In this work, we propose a topology-aware semantic infrared image colorization method, which includes a Topology-aware Semantic Constraint Module (TSCM) and a Discriminator-guided Attention Sampling Strategy (DASS). Although infrared and color images differ spectrally, they convey the same scene structure. TSCM shares adjacency matrices between specific patches of infrared and color images, utilizing graph neural networks to capture node feature representations across different domain graphs. By maximizing mutual information through contrastive loss, TSCM ensures semantic consistency in topology and preserves structural information. DASS selects the most informative patches during training, guiding the generator to maintain both content consistency and topological semantic consistency. Finally, experiments based on the KAIST, FLIR, and M3FD datasets demonstrate that our method outperforms state-of-the-art approaches in colorization.