On The Improving Algorithm for Stitching Images

Abstract

Image stitching is a fundamental problem in computer vision, enabling the creation of panoramas by aligning and blending multiple overlapping images. While traditional blending methods such as averaging or feathering are computationally efficient, they often produce visible artifacts including ghosting, seams, and intensity discontinuities, especially under exposure variations or geometric misalignments. Advanced techniques like multi-band or Poisson blending improve quality but are computationally expensive, limiting real-time applicability. To address these challenges, we propose a stitching framework that integrates rotation-robust geometric alignment with an optimized distance-based alpha blending strategy. By computing per-pixel distance maps from valid image regions and adaptively weighting contributions, the method achieves smoother transitions, reduced artifacts, and improved structural consistency. Extensive experiments demonstrate that the proposed approach outperforms classical and advanced blending techniques in both visual quality and computational efficiency, making it suitable for offline panorama generation as well as real-time video stitching applications.