Curvature Regularization for Non-Line-of-Sight Imaging From Under-Sampled Data

Area of Science:

• Computational imaging
• Inverse problems
• Photonics

Background:

• Non-line-of-sight (NLOS) imaging reconstructs hidden 3D scenes using time-of-flight photon data.
• Under-sampled scanning enables faster imaging but leads to ill-posed reconstruction problems sensitive to noise and distortions.

Purpose of the Study:

• To propose novel non-line-of-sight (NLOS) reconstruction models incorporating curvature regularization.
• To develop efficient optimization algorithms for these models, suitable for GPU implementation.

Main Methods:

• Introduced object-domain and dual (signal and object)-domain curvature regularization models.
• Developed optimization algorithms using the alternating direction method of multipliers (ADMM) with backtracking stepsize.
• Implemented solvers on Graphics Processing Units (GPUs) for accelerated computation.

Main Results:

• Achieved state-of-the-art performance on synthetic and real datasets, particularly in compressed sensing scenarios.
• Demonstrated superior effectiveness among iterative methods in balancing reconstruction quality and computational time.
• Validated the robustness of curvature regularization for NLOS imaging.

Conclusions:

• The proposed curvature regularization models and ADMM-based algorithms significantly improve NLOS reconstruction.
• GPU acceleration makes the method highly effective for practical, time-sensitive imaging applications.
• This work advances the field of computational imaging for complex, occluded environments.