Image stitching aims to composite images from different viewpoints into abroader field of view while ensuring precise alignment within the same perspective. However, significant parallax arise due to non-planar scenes and camera translations, presents challenges for accurately capturing the transformation relationship between two images using global or local homography methods. In this paper, we propose a novel unsupervised deep image stitching framework that combines global homography warping with per-pixel warping based on dense flow fields, to aim to achieve progressive refinement in image alignment from coarse to fine. In the fine alignment stage, an edge-guidance module is introduced to facilitate learning of edge-aware attention representations, enhancing the network’s ability to capture and understand critical boundaries and contours in images. Furthermore, to achieve more precise alignment, we design a flow prediction module incorporating scale softmax. This module suppresses weak correlations and enhances strong correlations, resulting in more accurate pixel displacement estimation. Extensive experimental validation demonstrates that our approach achieves superior accuracy compared to state-of-the-art solutions.