Abstract
Terahertz imaging offers significant potential in areas such as non-destructive testing, security screening, and medical diagnostics. However, due to the immature development of terahertz imaging devices, the field of view remains limited, making it challenging to capture complete target information in a single acquisition. While image stitching techniques can effectively expand the field of view, traditional methods encounter substantial limitations when applied to terahertz images, including low resolution, limited texture features, and inconsistencies arising from parallax. To address these challenges, particularly the parallax inconsistencies in low-resolution terahertz image stitching, we propose an Unsupervised Disparity-Tolerant Terahertz Image Stitching algorithm (UDTATIS). Our approach introduces targeted optimizations for two critical stages: geometric distortion correction and image feature fusion. Specifically, we design a feature extractor and an effective point discrimination mechanism based on the EfficientLOFTR architecture, significantly enhancing feature matching accuracy and robustness. Additionally, we introduce a continuity constraint to ensure the spatial continuity of matched points, thereby mitigating geometric distortions. Furthermore, we develop an improved conditional diffusion model that integrates multi-scale feature fusion with adaptive normalization, refining the transition effects along stitching boundaries. Compared to existing methods, UDTATIS demonstrates superior performance in handling terahertz images characterized by low resolution, limited textures, and parallax, achieving seamless image fusion while maintaining geometric consistency. Extensive quantitative and qualitative evaluations validate that UDTATIS outperforms state-of-the-art stitching algorithms, especially in complex scenes, delivering enhanced visual coherence and structural integrity. Project page: https://github.com/snow-wind-001/UDTATIS .