Deep Image Stitching for Unmanned Aerial Vehicles
Abstract
In the field of image stitching, existing techniques such as APAP, AANAP, ELA, and SPW utilize homography transformation and thin plate spline (TPS) for image warping. These methods achieve good results in ground‐level image stitching but struggle with the large parallax and complex feature scenes encountered in drone imagery. We propose an image stitching algorithm based on adaptive feature extraction and global feature enhancement module (AFGF), specifically designed for drone photography to address the challenges of stitching complex low‐altitude images. Our method integrates an adaptive feature extraction module and a global feature enhancement module. The adaptive feature extraction module effectively handles significant parallax caused by the rapid rotation and flight attitude changes of drones. The global feature enhancement module processes the warped images obtained from feature extraction, enabling robust global feature matching to seamlessly fuse the stitched images. Finally, we tested our method on five different low‐altitude drone image scenarios and achieved superior results compared to existing approaches. Extensive experiments validate the effectiveness of the proposed method in complex real‐ world scans, outperforming state‐of‐the‐art solutions by a significant margin.
