Abstract
While researches on model-based blind single image super-resolution (SISR) have achieved tremendous successes recently, most of them do not consider the image degradation sufficiently. Firstly, they always assume image noise obeys an independent and identically distributed (i.i.d.) Gaussian or Laplacian distribution, which largely underestimates the complexity of real noise. Secondly, previous commonly-used kernel priors (e.g., normalization, sparsity) are not effective enough to guarantee a rational kernel solution, and thus degenerates the performance of subsequent SISR task. To address the above issues, this paper proposes a model-based blind SISR method under the probabilistic framework, which elaborately models image degradation from the perspectives of noise and blur kernel. Specifically, instead of the traditional i.i.d. noise assumption, a patch-based non-i.i.d. noise model is proposed to tackle the complicated real noise, expecting to increase the degrees of freedom of the model for noise representation. As for the blur kernel, we novelly construct a concise yet effective kernel generator, and plug it into the proposed blind SISR method as an explicit kernel prior (EKP). To solve the proposed model, a theoretically grounded Monte Carlo EM algorithm is specifically designed. Comprehensive experiments demonstrate the superiority of our method over current state-of-the-arts on synthetic and real datasets.
Introduction
Blind Single Image Super-Resolution (SISR) aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) one where the specific degradation process—the blur kernel and the noise characteristics—is unknown. This task is particularly challenging in real-world scenarios where camera sensor noise is signal-dependent and spatially varying, contradicting the common assumption of independent and identically distributed (i.i.d.) Gaussian noise.
Existing methods often fail because they oversimplify noise modeling or use ineffective kernel priors that lead to unstable estimations. To address these flaws, the authors propose BSRDM (Blind Super-resolution with elaborate Degradation Modeling), a probabilistic framework that introduces a patch-based non-i.i.d. noise model and an Explicit Kernel Prior (EKP) to accurately capture real-world degradation.
Methodology
BSRDM models the degradation process by combining a deep generator for the HR image with explicit modeling of the blur kernel and noise.
Patch-based Non-i.i.d. Noise Model
Instead of a single global noise parameter, BSRDM treats the image as a collection of highly overlapped $p \times p$ patches. Each patch has its own variance parameter, allowing the model to adapt to the complex, non-uniform noise found in real camera sensors. This increased degree of freedom enables the model to represent noise more accurately than traditional i.i.d. assumptions.
Explicit Kernel Prior (EKP)
Rather than training a separate network to predict blur, the authors derive a concise, differentiable mathematical generator for anisotropic Gaussian kernels. This generator is controlled by a tiny $2 \times 2$ matrix (via Cholesky decomposition), ensuring the estimated kernel is always mathematically valid and physically plausible. This explicit constraint outperforms heuristic priors and avoids the instability of black-box neural networks.
Monte Carlo EM Algorithm
The model is solved using a theoretically grounded Monte Carlo Expectation-Maximization (EM) algorithm. In the E-Step, the latent HR image is sampled using Langevin dynamics. In the M-Step, the network parameters, the blur kernel, and the noise variance parameters are updated. For the image prior, the authors use a "Tiny Hourglass" architecture, which reduces the parameter count and speeds up inference compared to standard Deep Image Prior (DIP) models.
Experiments and Results
The researchers evaluated BSRDM on standard benchmarks (Set14, DIV2K100) and real-world datasets (RealSRSet).
Key Results
- Superior Accuracy: On the Set14 dataset (Scale x3) with complex camera noise, BSRDM achieved a PSNR of 26.24 dB, significantly outperforming the previous state-of-the-art DIPFKP (25.31 dB).
- Efficiency and Speed: BSRDM is remarkably lean, using only 762K parameters (compared to ~2.4M in competitors). It is also faster, completing a 4x upscale in roughly 190 seconds, whereas DIPFKP takes 333 seconds.
- Qualitative Quality: Visual results show that BSRDM produces much sharper edges and cleaner textures by successfully removing sensor noise that leaves other methods looking blurry or grainy.
Conclusion
BSRDM demonstrates that elaborate degradation modeling—specifically through patch-based noise representation and explicit kernel constraints—is essential for effective blind SISR. While the model is a conservative restorer that does not "hallucinate" details, its ability to generalize to unseen, complex degradations makes it a powerful tool for real-world image restoration. Future work will focus on developing even more powerful image priors to recover finer high-frequency textures.
Summary updated