Stable and performance-enhanced rectangling for image stitching using diffusion models

The panorama obtained from existing image stitching techniques often exhibit irregular boundaries, which affects the visual quality and usability. This paper proposes a novel diffusion-based approach that addresses the challenging task of image rectangling without compromising content integrity. This paper introduce two essential components: a Stable Motion Diffusion Model (SMDM) for generating initial rectangular transformation and a Content Diffusion Model (CDM) for refining details and enhancing the image quality. Additionally, we propose several architectural improvements, including a Selective Kernel Connection (SKC) for efficiently fusing features in skip connections and a Synergistic Sampling Block (SSBlock) for efficient downsampling. Extensive experiments on the Deep Image Rectangling Dataset (DIR-D) demonstrate that the model outperforms the state-of-the-art methods, achieving superior results in both numerical and visual quality.