Enhancing thermal infrared image colorization through reference-driven and contrastive learning approaches

Thermal infrared image colorization remains challenging due to limitations in existing methods, such as insufficient detail preservation and inaccurate color rendering. This paper presents a novel colorization approach that leverages reference images and contrastive learning to address these issues. Our model employs a dual-encoder generator architecture, allowing for detailed feature extraction from both infrared and reference images to enable precise color transfer. Key modules, including the Multi-Receptive Field Feature Integration Module (MFIM) and Channel–Spatial Feature Enhancement Module (CSFEM), enhance feature extraction and integration, while the Improved Stop-Gradient Attention Module (ISGA) ensures accurate feature alignment. A composite loss function combining adversarial, perceptual, and contrastive losses further refines the model’s output. Experimental results on benchmark datasets show that this method significantly improves colorization quality, generating visually realistic and detailed images, thus advancing applications in post-processing, object detection, and scene analysis within the infrared domain.