Wide Parallax Image Stitching With Balanced Alignment-Naturalness

Abstract

Parallax is a crucial factor that affects the quality of image stitching, particularly in complex scene images with wide camera baselines. Achieving precise alignment while maintaining natural visual aesthetics is a challenging task. Traditional stitching methods use fixed-grid partitioning and uniform local projective transformations, which can lead to distortion and nonlinear stretching in scenes with varying depths and orientations. This article proposes a novel approach to wide parallax image stitching that balances alignment precision with visual naturalness. Our method employs superpixel segmentation to partition images into regions with consistent local projective content. Additionally, local projective matrices derived from the Multi-GS algorithm (more efficient RANSAC) are used to transform inliers, thereby enhancing the accuracy of each superpixel block’s projective fitting, innovatively utilizing the guided local projective transformation algorithm. Furthermore, global similarity transformation and the linearization of local projective matrices are utilized to adjust overall distortion, while the combination of local projective transformation with global similarity is employed to achieve a more natural result. Experimental results show that our proposed algorithm outperforms state-of-the-art methods in improving alignment accuracy and the natural visual quality in scenarios with wide parallax. The RMSE and straight-line preservation improved by about 1% and 2% compared to LPC, while external expansion distortion and angular distortion improved by about 20% and 16% compared to UDIS-D.

Keywords

• Image stitching
• parallax
• superpixel
• precise alignment
• visual naturalness