Abstract:Atmospheric turbulence introduces severe spatial and geometric distortions, challenging traditional image restoration methods. We propose the Probabilistic Prior Turbulence Removal Network (PPTRN), which combines probabilistic diffusion-based prior modeling with Transformer-driven feature extraction to address this issue. PPTRN employs a two-stage approach: first, a latent encoder and Transformer are jointly trained on clear images to establish robust feature representations. Then, a Denoising Diffusion Probabilistic Model (DDPM) models prior distributions over latent vectors, guiding the Transformer in capturing diverse feature variations essential for restoration. A key innovation in PPTRN is the Probabilistic Prior Driven Cross Attention mechanism, which integrates the DDPM-generated prior with feature embeddings to reduce artifacts and enhance spatial coherence. Extensive experiments validate that PPTRN significantly improves restoration quality on turbulence-degraded images, setting a new benchmark in clarity and structural fidelity.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the serious spatial and geometric distortion problems encountered when imaging through atmospheric turbulence, which pose challenges to traditional image restoration methods. Specifically: 1. **Image degradation caused by atmospheric turbulence**: Atmospheric turbulence can lead to severe spatial and geometric distortions, causing a significant decline in image quality. This phenomenon is particularly common in fields such as remote monitoring, astronomy, and remote sensing, seriously affecting the reliability of high - fidelity image analysis. 2. **Limitations of existing methods**: - **Traditional computational methods**: They rely on simplified models and are not suitable for dynamic conditions. - **Convolutional Neural Network (CNN)**: It is difficult to capture long - distance dependencies, resulting in over - smoothed output images that lack detail and structural consistency. - **Transformer model**: Although it can capture long - distance dependencies, it faces challenges when dealing with highly uncertain and multi - modal turbulence - degraded images. 3. **Handling of high uncertainty and multi - modal distortion**: Atmospheric turbulence requires probabilistic modeling to handle multiple possible reconstruction results and avoid over - smoothing. At the same time, maintaining spatial coherence and detail integrity is crucial. ### Proposed solutions To solve the above problems, the authors propose the Probabilistic Prior Turbulence Removal Network (PPTRN) based on the diffusion model - driven attention mechanism with probabilistic prior. PPTRN combines the following key technologies: - **Latent Encoder**: Used to generate a compact representation of clear - image features. - **Denoising Diffusion Probabilistic Model (DDPM)**: Used to model the probabilistic prior distribution in the latent variable space and capture the multi - modal characteristics of clear images. - **Transformer architecture**: Used to extract detailed feature representations and perform restoration by combining the prior information generated by DDPM. - **Probabilistic Prior Driven Cross - Attention (PPDA)**: Fuses the prior information generated by DDPM with feature embeddings to reduce artifacts and enhance spatial coherence. Through the combination of these technologies, PPTRN can more effectively handle complex atmospheric turbulence distortion, restore high - quality images, and improve the clarity and structural fidelity of images. ### Main contributions 1. **Proposed a new framework (PPTRN)**, which combines probabilistic prior modeling with Transformer feature extraction for restoring turbulence - degraded images. 2. **Introduced the Probabilistic Prior Driven Cross - Attention mechanism**, which improves detail preservation and spatial coherence. 3. **Adopted a two - stage training strategy**, which balances structural consistency and detail preservation and captures the multi - modal features of turbulence - affected images. These innovations make PPTRN perform excellently in restoring turbulence - degraded images and improve the ability to handle uncertainty in visual restoration.

Probabilistic Prior Driven Attention Mechanism Based on Diffusion Model for Imaging Through Atmospheric Turbulence

Physics-Driven Turbulence Image Restoration with Stochastic Refinement

Effective Framework for Space Target Detection Through Atmospheric Turbulence

Spatio-Temporal Turbulence Mitigation: A Translational Perspective

Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise

Imaging Through the Atmosphere Using Turbulence Mitigation Transformer

Generative Diffusion Prior for Unified Image Restoration and Enhancement

Diffusion Models Without Attention

DR-DiT: Image Deraining Using Diffusion Model with Transformer

Adaptive Semantic-Enhanced Denoising Diffusion Probabilistic Model for Remote Sensing Image Super-Resolution

Diffusion Posterior Proximal Sampling for Image Restoration

A Unified Conditional Framework for Diffusion-based Image Restoration

Reconstruct-and-Generate Diffusion Model for Detail-Preserving Image Denoising

Distortion-driven Turbulence Effect Removal using Variational Model.

Restoring Vision in Adverse Weather Conditions With Patch-Based Denoising Diffusion Models

ReviveDiff: A Universal Diffusion Model for Restoring Images in Adverse Weather Conditions

ConVRT: Consistent Video Restoration Through Turbulence with Test-time Optimization of Neural Video Representations

Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration