Unsupervised Image Stitching: Image Alignment Method Based on Gradual Refinement Network

Image alignment is a critical step in the image stitching process. Traditional image alignment methods typically use uniform grid transformations or homography transformations to achieve geometric alignment. However, these methods have limitations when dealing with complex scenes, geometric distortions, and large viewpoint changes, making it difficult to preserve image details and local consistency. To address these issues, this paper proposes an image alignment method based on a Gradual Refinement Network. Specifically, a grid offset prediction network (Gradual Refinement Network) is designed for image alignment, which enhances sensitivity to detail changes through Haar wavelet downsampling, introduces spatial and channel collaborative attention to strengthen local information extraction, and employs a coarse-to-fine strategy to achieve precise image alignment. A loss function suitable for grid offset prediction is proposed to optimize the grid offset. Experimental results show that the proposed method surpasses existing image alignment techniques in handling complex geometric deformations and large viewpoint changes.