D²Turb: New AI clears blurry long-distance views by modeling scene depth

Atmospheric turbulence distorts long-range imagery with spatially varying blur and non-rigid geometric warping, a classic ill-posed problem that existing end-to-end deep learning approaches struggle to solve. These methods are typically trained on flat-field simulations and fail to balance fine texture recovery with accurate geometric rectification. D²Turb addresses this by first introducing a Depth-Aware Turbulence Synthesis protocol that incorporates scene depth into the phase-to-space formulation, generating physically realistic, depth-dependent degradations. This simulation engine also provides a crucial intermediate tilt supervision signal, enabling the model to learn disentangled representations of blur and distortion.

The framework then decomposes the restoration task into two interactive stages: texture deblurring and geometric rectification. The deblurring stage uses a dedicated backbone to recover fine-grained details while deliberately preserving geometric distortion for the subsequent rectification stage. To prevent information fragmentation typical in cascaded architectures, D²Turb introduces an Adaptive Structural Prior Injection (ASPI) mechanism that dynamically transfers deep structural representations from the deblurring module to guide dense flow prediction for spatial unwarping. Extensive evaluations on both synthetic benchmarks and real-world turbulence datasets demonstrate consistent state-of-the-art results in both texture fidelity and geometric accuracy.