Unsupervised Deep Image Stitching: Reconstructing Stitched Features to Images.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Traditional image stitching methods struggle with low-feature or low-resolution images.
• Supervised learning for image stitching is hindered by the scarcity of labeled data.

Purpose of the Study:

• To propose an unsupervised deep image stitching framework addressing limitations of existing methods.
• To develop a robust solution for stitching images in challenging scenarios like large-baseline scenes.

Main Methods:

• A two-stage framework: unsupervised coarse alignment and unsupervised reconstruction.
• Utilizing an ablation-based loss and transformer layer for alignment.
• Employing a dual-branch (low- and high-resolution) network for artifact elimination and resolution enhancement.

Main Results:

• The proposed method demonstrates superiority over state-of-the-art solutions in extensive experiments.
• Achieved competitive and preferred image stitching quality compared to supervised methods.
• Introduced a new benchmark dataset for unsupervised deep image stitching.

Conclusions:

• The unsupervised deep image stitching framework effectively handles challenging image pairs.
• The method offers a viable alternative to supervised approaches, especially when labeled data is unavailable.
• The introduced dataset and framework advance research in unsupervised image stitching.