Perceptual Quality Assessment of Real-World Noise in Ultra-High-Definition Images

Accurately assessing the perceptual quality of ultra-high-definition (UHD) images degraded by real-world noise remains a formidable challenge in computer vision. Existing Blind Image Quality Assessment (BIQA) methods, typically tailored for synthetic distortions and low-resolution inputs, often fail to generalize to these complex real-world scenarios. To bridge this gap, we first establish a real-world Ultra-High-Definition Noisy Image Quality (UNIQ-5k) dataset, consisting of 4,759 UHD images with fine-grained perceptual annotations. Leveraging this foundation, we propose PINE, a novel noise-oriented BIQA method. To tackle the computational burden of UHD inputs, we adopt a Grid-based Fragment Sampling Strategy (GFSS) as an efficient preprocessing component, which extracts representative local features while preserving global semantic topology. Subsequently, PINE incorporates a dual-branch architecture to simulate the noise visibility characteristics of the Human Visual System (HVS): the Perceptual Spatial Attention Module (PSAM) exploits the texture masking effect to prioritize noise in smooth regions; the Perceptual Frequency Enhancement Module (PFEM) utilizes a Modified Noise-Aware Contrast Sensitivity Function (CSF) to explicitly emphasize medium-to-high-frequency noise components. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the UNIQ-5k dataset and high rank consistency with PSNR and SSIM across the evaluated SIDD denoising results, showing that it may serve as an auxiliary no-reference indicator for restoration-oriented comparisons. The codes and models will be available at https://github.com/zijieweng123/PINE.git