Robust Two-Stage Image Stitching via Multi-Scale Homography Regression

Image stitching is a widely used and practical computer vision technique, which aims to generate images with a wide field of view. Traditional feature-based methods are heavily relying on the quality of geometric features, often showing poor performance in low-texture scenarios. Recently, deep learning-based approaches have demonstrated significant advantages in feature detection capability and robustness. Compared to traditional methods, deep learning-based stitching methods adaptively learn deep semantic features through powerful feature extraction networks, but they still face challenges in handling details of local areas. In this paper, we propose a two-stage supervised framework for image stitching. In the first stage, we design a supervised homography network that employs hybrid attention convolutional layers to extract multi-scale features and predicts homography transformations through multi-stage regression. In the second stage, we designed an attention-feature reconstruction module to better learn seam information and eliminate artifacts present in coarse-aligned images. Experimental results demonstrate that our approach outperforms existing homography prediction and image stitching methods both qualitatively and quantitatively, while also show better performance in local details processing.