TRACE method stabilizes image reconstruction with trajectory constraints
Training-free framework couples adjacent states to improve reconstruction quality across tasks.
A new approach called TRACE (TRAjectory-Constrained rEconstruction) tackles a key weakness in diffusion-based and iterative methods for imaging inverse problems. While these methods produce a trajectory of intermediate estimates, they rarely regulate transitions between steps. TRACE introduces explicit temporal coupling between consecutive states, interpretable as a sequence of proximal updates. Since exact proximal updates are intractable, the authors use a neural mapping approximation, yielding a diffusion-like process with stable, bounded trajectory variation even with untrained networks.
Tests on linear and nonlinear image reconstruction tasks (e.g., denoising, inpainting) show TRACE consistently improves output quality. Ablation studies confirm that temporal coupling directly controls transitions along the reconstruction path. The framework requires no training—a major advantage for practical deployment—and provides a stability analysis proving that coupling bounds trajectory variation. With 20 pages and 10 figures, the paper provides a rigorous foundation for trajectory-constrained reconstruction, potentially benefiting medical imaging, computational photography, and scientific imaging where reliable, artifact-free results are critical.
- TRACE introduces explicit temporal coupling between consecutive reconstruction states, unlike existing methods that ignore trajectory transitions.
- The training-free framework uses a neural mapping to approximate intractable proximal updates, enabling stable diffusion-like reconstruction.
- Experiments on linear and nonlinear tasks (including denoising and inpainting) demonstrate improved quality, with stability analysis proving bounded trajectory variation.
Why It Matters
Enables stable, high-quality image reconstruction without training—critical for medical and scientific imaging applications.