Content-Seam-Preserving Multi-Alignment Network for Visual-Sensor-Based Image Stitching.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Image stitching is crucial for creating panoramic views in virtual and augmented reality.
• Current deep learning methods often result in alignment distortions due to single homography estimation.
• There is a need for methods that maintain content consistency and prevent seam artifacts.

Purpose of the Study:

• To propose a novel Content-Seam-Preserving Multi-Alignment Network (CSPM-Net) for visual-sensor-based image stitching.
• To address limitations of existing methods by preserving image content and avoiding seam distortions.
• To enhance the quality of stitched panoramic images for immersive applications.

Main Methods:

• Developed a content-preserving deep homography estimation for initial image pair alignment.
• Implemented edge-assisted mesh warping incorporating edge information to eliminate seam artifacts.
• Introduced a content consistency loss and a seam smoothness loss for accurate final image prediction.

Main Results:

• The CSPM-Net effectively reduces content inconsistency and seam distortions.
• Experimental results show superior stitching performance compared to state-of-the-art methods.
• The method preserves the geometric structure of overlapping regions and eliminates edge distortions.

Conclusions:

• The proposed CSPM-Net offers a robust solution for visual-sensor-based image stitching.
• This approach significantly improves the naturalness and quality of panoramic images.
• CSPM-Net outperforms existing methods in preserving content and minimizing visual artifacts.