Region-Based Self-Segmentation Guided Diffusion Model for Thermal Infrared to Pseudo-Color Visible Light Image Conversion

Unlike traditional visible light cameras, thermal infrared cameras can be used to conduct clear reconnaissance and observations during both day and night. However, because of the imaging principle of thermal infrared cameras, thermal infrared images lack color features and appropriate detail information for human observation. This paper introduces RSDM, a novel Region-aware Semantic-guided Diffusion Model, for converting thermal infrared images into perceptually accurate pseudo-color visible light images. RSDM addresses the limitations of existing methods by incorporating a two-stage process: 1) non-uniformity correction to remove sensor artifacts; and 2) unique SEM-guided diffusion model for image generation. This approach leverages pre-trained diffusion models in conjunction with region segmentation techniques to achieve superior performance, particularly in preserving fine details and effectively handling small targets. We systematically evaluated the proposed RSDM method on the FLIR, M3FD, and self-collected datasets. The experimental results indicate that our approach yields significant enhancements in image quality, detail preservation, and object recognition performance compared to current state-of-the-art methods.