Parallax tolerant image stitching based on concise registration

Large parallax image stitching has long been a major challenge due to its high complexity and numerous interference factors. Large baselines and moving objects during imaging often lead to distortion, blurring, ghosting, and object cutting or duplication in stitched results. This paper focuses on two major challenges in natural image stitching: (1) significant parallax induced by viewpoint changes; and (2) object cutting or duplication caused by moving objects. Since introducing or optimizing algorithms in a single stage is insufficient to address the composite challenges of large parallax, we design targeted solutions for each of the three stitching stages. Specifically, in the registration stage, we propose a novel registration algorithm, FMSS, to obtain multi-plane and multi-depth registration pairs, while identifying disparity or high-risk regions. It achieves simplicity and speed with minimal accuracy loss compared to traditional methods. In the seam finding stage, based on the disparity or high-risk regions obtained earlier, we employ SAM to extract these object regions and improve the energy function design of the graph-cut MRF, incorporating superior region terms and duplication avoidance terms to guide the seam to pass through less error-prone areas. Finally, in the blending stage, we adopt the commonly used alpha blending technique. We validate our method on a challenging dataset, and the results demonstrate that it outperforms current state-of-the-art approaches. The code will be available at https://github.com/xingfuhoumei/Parallax-Tolerant-Image-Stitching-Based-on-LCR-and-OD.