DIPLI Improves Blind Astronomical Image Restoration

Suraj Singh, Anastasia Batsheva, Oleg Y. Rogov, and Ahmed Bouridane present DIPLI (Deep Image Prior Lucky Imaging), a method for blind astronomical image restoration that moves the Deep Image Prior approach from single-frame fitting to multi-frame training. The framework integrates Back Projection for multi-frame signal fusion, optical-flow-based alignment, and stochastic sampling to reduce overfitting, and it reports consistent gains across perceptual and fidelity metrics over classical and learned baselines.

DIPLI replaces per-image deterministic DIP optimization with three coordinated components. Key elements are:

• •Multi-frame training using Back Projection to enforce consistency across short-exposure frames.
• •Optical flow estimation with TVNet to align turbulent, misregistered frames before fusion.
• •Unbiased Monte Carlo estimation achieved by Langevin dynamics to inject stochasticity and prevent overfitting and artifacts.

The authors evaluate against Lucky Imaging, vanilla Deep Image Prior, the transformer-based RVRT, and the diffusion-based DiffIR2VR-Zero. Metrics reported include SSIM, PSNR, LPIPS, and DISTS, with DIPLI outperforming baselines on synthetic datasets and showing robust reconstructions on real telescope data while using fewer input frames than Lucky Imaging.

Deep Image Prior is attractive for astrophotography because it requires no external training data, but it is prone to overfitting and hallucination on noise-limited, low-frame settings. DIPLI addresses those failure modes by enforcing cross-frame consistency and introducing principled stochastic estimation. This aligns with broader trends toward hybrid methods that combine classical optics-aware alignment, physics-inspired operators, and modern neural priors. For practitioners working with resolved, high-contrast targets, DIPLI offers a lower-data-cost alternative to fully supervised models and can be integrated into existing post-processing pipelines.

Validate DIPLI on more diverse atmospheric conditions, extended targets, and real-time pipelines. Pay attention to computational cost of Langevin dynamics sampling and sensitivity to optical flow errors from TVNet when frames have severe distortions.