UHD Image Dehazing via anDehazeFormer with Atmospheric-aware KV Cache

2025-05-20

AI Summary

To address the bottlenecks of slow training convergence and high GPU memory consumption in ultra-high-resolution (4K/8K) image dehazing, this paper proposes a lightweight and efficient vision Transformer architecture. Methodologically: (i) an nGPT-inspired adaptive normalization mechanism is introduced to significantly accelerate training convergence; (ii) a physics-guided key-value (KV) caching strategy, grounded in the atmospheric scattering model, is designed to dynamically preserve haze-relevant features with high fidelity; and (iii) gradient attribution is integrated to enable interpretable analysis. Experiments demonstrate that the proposed method achieves real-time inference at 50 FPS on an RTX 4090, accelerates training by 5×, and reduces GPU memory usage by ~40%, while maintaining state-of-the-art dehazing quality on standard benchmarks including SOTS and RESIDE—enabling efficient and robust restoration of high-resolution images.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Hardware-aware ML
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Security and Privacy: Data transparency and provenance
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

In this paper, we propose an efficient visual transformer framework for ultra-high-definition (UHD) image dehazing that addresses the key challenges of slow training speed and high memory consumption for existing methods. Our approach introduces two key innovations: 1) an extbf{a}daptive extbf{n}ormalization mechanism inspired by the nGPT architecture that enables ultra-fast and stable training with a network with a restricted range of parameter expressions; and 2) we devise an atmospheric scattering-aware KV caching mechanism that dynamically optimizes feature preservation based on the physical haze formation model. The proposed architecture improves the training convergence speed by extbf{5 $ imes$} while reducing memory overhead, enabling real-time processing of 50 high-resolution images per second on an RTX4090 GPU. Experimental results show that our approach maintains state-of-the-art dehazing quality while significantly improving computational efficiency for 4K/8K image restoration tasks. Furthermore, we provide a new dehazing image interpretable method with the help of an integrated gradient attribution map. Our code can be found here: https://anonymous.4open.science/r/anDehazeFormer-632E/README.md.

Problem

• Improve training speed for UHD image dehazing
• Reduce memory consumption in dehazing methods
• Enhance computational efficiency for 4K/8K restoration

Innovation

• Adaptive normalization for fast stable training
• Atmospheric-aware KV cache optimizes features
• Real-time 50-image processing on RTX4090