Abstract
Ultra-High-Definition (UHD) image deblurring is increasingly critical as 4K displays become standard, yet it remains computationally prohibitive for most state-of-the-art models. This paper introduces SeMIR, a framework designed for extreme efficiency and robustness. SeMIR addresses the limitations of standard restoration networks by "softening" semantic priors from foundation models to account for uncertainty and utilizing an Implicit Feature Processor (IFP) to recover high-frequency details lost during downsampling. Our model achieves a state-of-the-art PSNR of 31.38 dB on the UHD$^2$B benchmark with an inference time of only 2ms per 4K image, significantly outperforming heavier Transformer-based architectures.
Introduction
The demand for UHD image restoration has surged, but two primary obstacles remain: the massive computational cost of processing 4K resolutions and the permanent loss of high-frequency information during network downsampling. While recent trends involve using priors from large foundation models like the Segment Anything Model (SAM), these approaches often suffer from "over-trusting" the prior; errors in the initial segmentation can lead to catastrophic failures in the restored output.
Methodology: The SeMIR Framework
SeMIR is built upon an efficiency-first architecture that integrates semantic guidance with implicit frequency constraints.
Uncertainty Semantic Prior
To mitigate the risks of incorrect semantic guidance, SeMIR introduces a "softening" factor ($\lambda$). Instead of treating raw segmentation maps as absolute truth, the model models the uncertainty of these labels. This allows the network to prioritize actual image content when the semantic prior is unreliable, ensuring robustness against noise in blurry regions.
Implicit Feature Processor (IFP)
To combat information loss during downsampling, we developed the IFeX (Implicit Feature Extraction) module. Inspired by Fourier analysis, IFeX decomposes signals into sinusoidal waves of varying frequencies. This allows the model to implicitly represent and reconstruct the fine textures and high-frequency details essential for UHD quality that are typically discarded by standard convolutional layers.
Datasets
To evaluate real-world performance, we introduced the UHD$^2$B dataset, consisting of 2,025 pairs of real UHD images and realistically synthesized blurry counterparts. We also utilized the 4KRD dataset. Quantitative analysis shows that UHD$^2$B achieves the lowest Fréchet Inception Distance (FID), indicating it is highly representative of real-world degradation.
Key Results
SeMIR establishes a new frontier in the trade-off between restoration quality and computational complexity.
- Efficiency: With only 9.7 million parameters, SeMIR is nearly three times leaner than Restormer (26.1M). It achieves a remarkable inference speed of 2ms for a 4K image.
- Performance: On the UHD$^2$B benchmark, SeMIR reached a PSNR of 31.38 dB, outperforming Uformer (28.48 dB) and Restormer (26.10 dB).
- Visual Fidelity: Qualitative assessments show superior restoration of fine textures, such as text on license plates and architectural details, avoiding the "watercolored" artifacts common in other methods.
Conclusion
SeMIR demonstrates that high-performance UHD deblurring does not require massive Transformer architectures. By intelligently handling semantic uncertainty and employing implicit frequency constraints, SeMIR provides a fast, robust solution for real-time applications. Future work will focus on integrating a lightweight segmentation head to create a fully end-to-end real-time pipeline.
Summary updated