Source-linked AI summary
SLER-IR: Spherical Layer-wise Expert Routing for All-in-One Image Restoration
Peng Shurui, Xin Lin, Shi Luo, Jincen Ou, Dizhe Zhang, Lu Qi, Truong Nguyen, Chao Ren
TL;DR
Unified image restoration must handle diverse degradations despite feature interference, insufficient expert specialization, and routing limitations. SLER-IR uses spherical layer-wise expert routing with hyperspherical degradation embeddings and global–local fusion, and reports consistent PSNR/SSIM improvements across three-task and five-task benchmarks.
Problem
Unified restoration remains challenging because diverse degradations create feature interference, insufficient expert specialization, and geometry-biased routing representations.
Method
SLER-IR dynamically selects parameter-independent experts layer by layer using hyperspherical degradation embeddings, contrastive learning, and global–local granularity fusion.
Results
SLER-IR consistently outperforms state-of-the-art methods in PSNR and SSIM across three-task and five-task benchmarks.
Takeaways & Limitations
The framework supports compositional inference paths and progressive expert specialization for diverse and composite degradations without increasing inference overhead.
Abstract
from arXiv · showhide
Image restoration under diverse degradations remains challenging for unified all-in-one frameworks due to feature interference and insufficient expert specialization. We propose SLER-IR, a spherical layer-wise expert routing framework that dynamically activates specialized experts across network layers. To ensure reliable routing, we introduce a Spherical Uniform Degradation Embedding with contrastive learning, which maps degradation representations onto a hypersphere to eliminate geometry bias in linear embedding spaces. In addition, a Global-Local Granularity Fusion (GLGF) module integrates global semantics and local degradation cues to address spatially non-uniform degradations and the train-test granularity gap. Experiments on three-task and five-task benchmarks demonstrate that SLER-IR achieves consistent improvements over state-of-the-art methods in both PSNR and SSIM. Code and models will be publicly released.
1 Introduction
All-in-one restoration seeks one model for diverse degradations, but existing feature-modulation and expert-based paradigms face interference, limited specialization, and routing challenges. SLER-IR addresses these issues with spherical layer-wise routing, degradation-aware embeddings, and global–local fusion, while reporting broad benchmark gains.
- Task-specific restoration models lack flexibility for complex or unseen degradations and require labor-intensive retraining for new scenarios.
- Feature-modulation methods can suffer shared-backbone interference, while expert architectures may restrict experts to local modules and underuse degradation priors.
- SLER-IR replaces backbone blocks with parameter-independent experts and dynamically activates them layer by layer as features evolve.
- Spherical degradation embedding with triplet-based contrastive learning mitigates linear-space class-distance bias and improves degradation discriminability for routing.
- GLGF combines global image context with local degradation cues to bridge patch-based training and full-resolution inference.
2 Related Works
Related work spans task-specific restoration, general-purpose models, prompt-based feature modulation, and architectural expert adaptation. These approaches motivate unified all-in-one restoration for multiple degradations without explicit task identification or separate fine-tuning.
- Single-task methods target known degradation types and often generalize poorly to unseen or mixed degradations.
- General-purpose models improve cross-task generality but still require retraining or task-specific adaptation for unseen or composite degradations.
- SLER-IR is organized around spherical layer-wise expert routing, similarity-based gating, hyperspherical degradation representation learning, and global–local granularity fusion.
- All-in-one restoration trains one unified model to handle multiple degradations without explicit task identification or separate fine-tuning.
- Prompt-based methods modulate features in a shared parameter space using degradation, language-guided, or content-decoupled representations.
3 Method
SLER-IR builds a unified restoration model with layer-wise expert routing, spherical degradation embeddings, and global-local fusion. These components support progressive specialization, geometry-aware branch selection, and restoration guidance for spatially varying degradations.
- Spherical Layer-wise Expert Routing: SLER-IR expands each encoder-decoder layer into three parameter-independent experts, forming a unified multi-expert backbone.The architecture provides distinct latent feature representations at every layer.
- Spherical Layer-wise Expert Routing: Independent routing across eight layers enables up to 6561 compositional inference paths for complex and composite degradations.Routing trajectories adapt progressively across layers rather than remaining fixed per task.
- Router Design and Similarity-based Expert Gating: The router extracts degradation cues, produces L2-normalized layer-wise vectors, and selects experts using cosine similarity to expert centers.L2 normalization places routing features on the unit hypersphere; probabilistic routing is used during router optimization.
- Router Design and Similarity-based Expert Gating: Stage I learns stable degradation-aware representations with probabilistic routing, while Stage II freezes the router and activates the maximum-probability expert at each layer.Fixed routing allows restoration refinement and progressive specialization as features evolve across network depth.
- Hyperspherical Degradation Representation Learning: Spherical uniform embedding maps degradation representations onto a unit hypersphere, reducing geometry-induced distance bias and improving branch selection through angular comparison.Triplet-constrained contrastive optimization encourages compact within-class and separated cross-class embeddings, yielding more reliable routing scores.
- Global–Local Granularity Fusion: GLGF aligns global content context with localized degradation evidence by constructing CSP and DSP maps and fusing them into a restoration prior map.This design targets spatially non-uniform degradations and the granularity mismatch between patch-based training and full-image inference.
4 Experiment
The experiments evaluate SLER-IR across three-task and five-task all-in-one restoration settings using PSNR and SSIM, quantitative comparisons, ablations, and visual results. SLER-IR consistently outperforms competing methods while improving restoration quality across diverse degradations.
- Experimental Setup: Evaluations cover three-task and five-task all-in-one settings using PSNR and SSIM across denoising, dehazing, deraining, deblurring, and low-light enhancement.The five-task datasets include BSD400 and WED for denoising, SOTS for dehazing, Rain100L for deraining, GoPro for deblurring, and LOL for low-light enhancement.
- Quantitative Evaluation: 33.14/0.922 in PSNR/SSIM is SLER-IR’s highest average performance in the three-task setting, exceeding MoCE-IR by 0.41 dB and 0.005, respectively.It also improves dehazing PSNR on SOTS by 1.66 dB and gains consistently across CBSD68 noise levels σ = 15, 25, 50.
- Quantitative Evaluation: 31.73/0.928 in PSNR/SSIM is SLER-IR’s highest overall average in the five-task setting, surpassing the previous approach by 1.15 dB and 0.009, respectively.Reported gains include +2.59 dB on SOTS dehazing, +1.22 dB on GoPro deblurring, and +0.96 dB on LOL low-light enhancement.
- Qualitative Evaluation: Visual comparisons show clearer textures and sharper edges for deblurring, cleaner colors for dehazing, more thorough rain removal, and preserved scene details under rainy conditions.For low-light enhancement, SLER-IR achieves balanced enhancement with well-preserved structures and natural illumination.
- Ablation Studies: Ablations report benefits from auxiliary classifier supervision and Global–Local Granularity Fusion, which combines global semantic cues with local degradation evidence.The loss-function ablation evaluates added degradation-aware supervision, while the GLGF ablation compares against a baseline without the module.
5 Conclusion
SLER-IR is presented as a spherical layer-wise expert routing framework for unified all-in-one image restoration. Its hyperspherical embedding, global–local guidance, and compositional inference paths support progressive specialization and are associated with stronger quantitative and visual results across multiple benchmarks.
- Conclusion: SLER-IR uses hyperspherical degradation embedding for geometry-consistent routing and global–local granularity fusion for spatially robust guidance.The framework is designed for progressive expert specialization under diverse degradations.
- Conclusion: The routing mechanism enables flexible compositional inference paths without increasing inference overhead.This supports adaptation to complex degradation scenarios within the unified restoration framework.
- Conclusion: Across multiple benchmarks, SLER-IR consistently outperforms state-of-the-art methods in quantitative metrics and visual quality.