Learnable explicit degradation model for blind hybrid-distorted image restoration

Real-world images captured by practical imaging pipelines (e.g., mobile cameras and consumer devices) are often degraded by unknown and hybrid distortions (e.g., blur, noise, and compression artifacts). Although neural network-based methods for hybrid-distorted image restoration have achieved remarkable success, they are trained on synthetic paired images and do not perform well on real-world target-device test images. To narrow this gap, researchers learn test-data distortions from unpaired distorted-clean images via unsupervised learning. However, modeling the entire degradation process with a single integrated network induces a high-dimensional, complex optimization space that is difficult to train; as a result, the learned degradations fail to faithfully capture real-world distortions. To address this issue, we propose a Learnable Explicit Degradation Model (LEDM), a modular degradation generator designed to learn test-data distortions under the unpaired setting. LEDM splits the degradation into three interpretable subspaces, each learned by a dedicated network. This explicit design makes the learning process more structured and helps alleviate optimization difficulty. The modular design also enables fine-grained design. The noise learning module considers whether the noise is correlated with signal strength and whether the noise itself is spatially correlated, leading to more realistic noise synthesis. Moreover, we incorporate physics-informed regularization losses to constrain the learned blur kernel and noise, thereby preventing the generation of unrealistic distorted images. By using Gaussian-distributed inputs, LEDM can generate more diverse degradations, helping reduce the degradation gap. Experimental results show that restoration models trained on LEDM-generated pairs achieve improved performance on target-device real-world datasets such as DPED and Real-Nikon, demonstrating the effectiveness of the proposed method for specific imaging pipelines.