Uncertain Semantics Meet Implicit Constraints: A New Frontier in Real-World UHD Image Deblurring

Image deblurring has always been a challenging problem in computer vision, especially when dealing with real-world images. Despite the significant progress made by current methods, they often struggle to handle Ultra-High-Definition (UHD) images due to computational resource limitations. In this paper, we propose a novel UHD image restoration model called SeMIR, which focuses on the UHD deblurring task and can run a full-resolution UHD image with a single GPU shader. The design of our approach is based on two well-established insights. First, existing image enhancement methods with the help of foundation models over-trust the decisions of foundation models, here we model uncertainty by modeling the decisions of foundation models to boost the performance of image deblurring. Second, explicit deep network inference tends to exhibit overfitting. Here, we design an Implicit Feature Processor (IFP) module, which constrains the model’s inference to accurately output clear images through global and local modeling. Meanwhile, we also contribute a new UHD image dataset, UHD2B, which comprises 2, 025 pairs of UHD images. The dataset includes a diverse range of scenes and environmental conditions and provides a standardized benchmark for evaluating and comparing different methods. By evaluating both quantitatively and qualitatively on two standard benchmarks, SeMIR’s efficient performance is remarkable, with an inference time of only 2 ms for each 4K image.