Technical Summary: Circular Quasiconformal Deturbulence (CQCD)

Problem Statement

Image restoration through inhomogeneous media, such as turbulent air or water, results in severe geometric distortions, blurring, and noise. A primary challenge in addressing this is the lack of clean, paired training data (distorted image vs. ground truth) required for supervised learning. Capturing both the degraded and pristine versions of a scene under identical conditions is often infeasible. Existing unsupervised approaches often struggle to generalize across different distortion types or introduce artifacts due to imperfect distortion modeling. Furthermore, many methods fail to ensure the invertibility of deformation mappings, which is critical for cycle-consistency approaches, leading to error accumulation and non-physical artifacts like folding or tearing.

Methodology

The authors propose Circular Quasi-Conformal Deturbulence (CQCD), an unsupervised framework that reconstructs distortion-free images from multiple turbulent frames using a circular self-consistency architecture. The method operates through a three-stage cycle:

1. Distortion-Free Image Estimation (Forward Pass):

   • Tight-Frame (TF) Encoding: Input distorted frames undergo a tight-frame transformation to extract multi-scale, direction-aware features. This explicitly separates horizontal and vertical frequency components, allowing the network to capture orientational deformation patterns more effectively than standard Fourier-based encodings.
   • Mapping Estimation: A neural network (Deformation Estimator, $N_{DE}$) predicts a geometric deformation field ($f_t$) to warp the distorted images into a corrected form.
   • Resampling: The distorted images are warped using the estimated field. This process mitigates geometric turbulence but may introduce interpolation blur.
   • Deblurring: A pixel-wise network (Blur Remover, $N_{BR}$) processes the resampled features to produce a sharp, restored image ($I^*$).

2. Distorted Image Estimation (Inverse Pass):

   • To enforce cycle consistency, the inverse of the estimated deformation mapping ($f_t^{-1}$) is applied to the restored image $I^*$.
   • This reconstructs an approximation of the original distorted input ($\bar{I}_t$).

3. Optimization and Regularization:

   • Loss Functions: The framework minimizes a reconstruction loss ($L_{rec}$) between the restored image and the warped inputs, and a distortion loss ($L_{dist}$) between the reconstructed distorted images and the original inputs.
   • Quasi-Conformal (QC) Regularization: A critical component is the integration of computational quasi-conformal geometry. The deformation fields are regularized using the Beltrami coefficient ($\mu$), which quantifies local conformality distortion. By minimizing the $L_2$ norm of the Beltrami coefficients for both forward and inverse mappings, the model encourages bijective, homeomorphic deformations. This ensures the mappings are invertible and prevents non-physical artifacts like folding, which is essential for the circular loop to function without error accumulation.

The total loss function for the deformation estimator combines fidelity terms with a QC regularization term weighted by $\lambda$, while the blur remover is optimized primarily on fidelity terms.

Key Contributions

1. Circular Turbulence Removal Framework: The authors propose a novel unsupervised architecture that utilizes multiple observations to iteratively refine restoration. The circular design promotes temporal and spatial consistency and acts as an implicit validation mechanism to prevent error accumulation.
2. Quasi-Conformal Guidance: The paper integrates computational quasi-conformal theory to guide deformation mappings toward homeomorphic and invertible behavior. This is mathematically grounded in the Beltrami coefficient, ensuring the forward and inverse mappings remain bijective, which is a prerequisite for the circular consistency constraint.
3. Tight-Frame Encoding: The introduction of tight-frame blocks enables the CNN to directly capture orientational deformation patterns (horizontal vs. vertical jitter) without learning fundamental frequency representations from scratch. This enhances the precision of deformation field estimation and subsequent blur correction.

Experimental Results

The effectiveness of CQCD was evaluated on both synthetic and real-world datasets involving water and atmospheric turbulence.

• Quantitative Performance: On synthetic underwater datasets (Jellyfish, Pterois, Tiles) with varying turbulence intensities (mild, medium, severe) and input frame counts (5, 10, 15), CQCD outperformed state-of-the-art unsupervised methods (NeRT, NDIR, DiffTemplate, TurbRecon) in both PSNR and SSIM metrics. The performance gap widened as turbulence severity increased.
• Visual Quality: In real-world scenarios, CQCD demonstrated superior ability to restore geometric structures (e.g., straight lines of bricks, railings) and fine details (e.g., text legibility, jellyfish tentacles) compared to competitors, which often exhibited residual warping, color bleeding, or ghosting.
• Deformation Field Accuracy: Comparisons with ground-truth deformation fields used in synthetic simulations showed that CQCD accurately estimated the underlying distortion patterns, validating its physical interpretability.
• Ablation Studies:
  • QC Regularization: Removing QC regularization ($\lambda=0$) led to non-invertible mappings, folding artifacts, and failure to correct distortions. An optimal weight ($\lambda=0.1$) balanced flexibility with stability.
  • Tight-Frame Encoding: Excluding the TF block resulted in significant performance drops across all turbulence levels, confirming its necessity for capturing directional deformation features.
  • Architecture: A shallow U-Net (depth 2) was found sufficient for deformation estimation, and a balanced training ratio between the blur remover and mapping estimator yielded the best results.

Significance and Claims

The paper claims that CQCD represents a significant advancement in unsupervised image restoration by addressing the critical issue of mapping invertibility through quasi-conformal geometry. The authors assert that their framework not only surpasses existing deturbulence techniques in restoration quality but also provides a physically interpretable estimation of deformation fields. The circular architecture, coupled with QC constraints, allows the model to effectively handle complex geometric distortions without paired training data, making it suitable for challenging real-world scenarios like underwater and atmospheric imaging. The work highlights that enforcing bijectivity is not merely an auxiliary constraint but a fundamental requirement for the stability and success of cycle-consistent restoration models.

Limitations

The authors acknowledge that the current framework is designed for static scenes. The pixel-wise correspondence assumption inherent in the circular consistency loss is violated by moving objects or global camera motion, which the model may misinterpret as geometric distortion. Additionally, the reliance on quasi-conformal regularization limits the method to bijective mappings, meaning it cannot handle scenes involving topology-changing deformations. Future work is suggested to explore motion disentanglement and alternative feature extraction methods like wavelet transforms.