An Unsupervised Image Stitching Framework via Joint Iterative Optimization of Deformation Estimation, Feature...

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Image stitching creates panoramic images but struggles with uncorrected radial lens distortion.
• Consumer cameras often lack built-in distortion correction, leading to alignment issues in standard stitching.

Purpose of the Study:

• To develop a unified framework for image stitching that integrates unsupervised radial distortion correction with feature registration and seam blending.
• To address limitations of conventional methods that assume pre-corrected imagery and degrade performance with uncorrected wide-angle or fisheye lenses.

Main Methods:

• A task-driven joint iterative optimization framework is proposed, modeling lens distortion parameters as learnable variables.
• A closed-loop optimization strategy refines distortion parameters, homography estimates, and seam paths iteratively.
• A calibration-free initial distortion estimation method uses image gradients and epipolar consistency.

Main Results:

• The proposed method achieves superior stitching fidelity, measured by PSNR and SSIM, on benchmarks with significant radial distortion.
• It demonstrates enhanced feature matching stability compared to distortion-agnostic and two-stage approaches.
• Ablation studies validate the effectiveness and synergistic contribution of each module in the joint optimization architecture.

Conclusions:

• The unified framework effectively handles radial distortion within the image stitching pipeline.
• Joint optimization of distortion correction and stitching leads to more accurate and visually appealing panoramic images.
• The method offers a robust solution for stitching images from cameras lacking intrinsic calibration or distortion correction.