Progressive alignment and interwoven composition network for image stitching

Abstract

As one of the fundamental tasks in computer graphics and image processing, image stitching aims to combine multiple images with overlapping regions to generate a high-quality naturalness panorama. Most deep learning based image stitching methods suffer from unsatisfactory performance, because they neglect the cooperation relationship and complementary information between reference image and target image. To address these issues, we propose a progressive alignment and interwoven composition network (PAIC-Net) to produce satisfactory panorama images, which learns the cooperation relationship by a progressive homography alignment module and captures the complementary information by an interwoven image composition module. Specifically, a progressive homography alignment module is presented to align the input images, which progressively warps the reference and target images by focusing more on the combination of self-features and cooperation features. Then, an interwoven image composition module is presented to seamlessly fuse aligned image pairs, where the complementary information of one-view is captured to guide another-view in an interweaved way. Finally, an alignment loss and a composition loss are introduced to reduce alignment distortions and enhance seam consistency of the final image stitching results. Experimental results on benchmark datasets demonstrate that PAIC-Net outperforms state-of-the-art image stitching methods both quantitatively and qualitatively.