Vol. 56, Issue 2, pp. 327-339 (2026)

Abstract

Image stitching is a key computational imaging technique for expanding the field of view of optical systems and generating high-resolution panoramas. However, eliminating ghosting artifacts and geometric distortions remains challenging under non-ideal imaging conditions involving large parallax or weak textures. To address these issues, this paper proposes a deep learning image stitching network based on large-kernel feature perception and adaptive confidence gating. First, ConvNeXt-Tiny is adopted as the feature extraction backbone to capture long-range geometric dependencies and global context. Second, an adaptive confidence gating mechanism dynamically suppresses erroneous mesh deformations in uncertain regions to ensure precise geometric alignment. Finally, an attention-guided fusion network optimizes feature alignment and suppresses artifacts. Experimental results on the UDIS-D dataset demonstrate that this method outperforms existing state-of-the-art approaches, achieving a 6.44% improvement in structural similarity index (SSIM) to 0.892 and a 0.13 increase in peak signal-to-noise ratio (PSNR) to 25.55. This approach effectively resolves parallax artifacts in image stitching and significantly enhances the structural consistency of panoramic images.