Loop-Mamba: A Loop Mamba with Degradation-Aware and Shared Memory for Old Photo Restoration

summary

The gist

The gist The proposed Loop-Mamba framework formulates old photo restoration as a progressive loop state evolution process, enabling continuous restoration through recurrent state propagation rather

In short

Loop-Mamba restores old photos by treating restoration as a progressive loop state evolution rather than repeated image reconstruction. It introduces a Semantic-Guided Degradation Estimator to model different types of damage and a Shared Structural Memory to maintain long-range structural consistency across iterations. This lightweight framework achieves superior quality and efficiency on old photo restoration benchmarks.

Key concepts

Progressive State Evolution: This approach frames image restoration as a series of steps where the image's state is continuously refined through iterative computation. Instead of trying to fix everything at once, the model updates its internal 'state' step-by-step, guided by information about how degraded the photo is at each stage.
Semantic-Guided Degradation Estimator (SGDE): SGDE explicitly models different types of damage in old photos. It predicts two things: a map showing where specific local degradations (like scratches or noise) are occurring, and a global score summarizing the overall degradation level. This information guides how the restoration state should evolve.
Shared Structural Memory (S$^2$M-Mamba): This mechanism maintains a persistent memory of the photo's structure throughout all restoration iterations. By sharing this memory, the model ensures that long-range structural details remain consistent and robust as it iteratively refines the image, preventing structural inconsistencies.
Multi-directional State Propagation: Instead of processing states sequentially, this strategy propagates them along multiple spatial directions simultaneously using a lightweight scanning method. This enriches the information captured in each state update by looking at the photo from several angles at once, improving long-range interaction modeling with minimal extra computation.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Federated Learning — explained in 17 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Interpretability — explained in 15 episodes
Quantum Fisher Information (QFI) — explained in 15 episodes
Sample Complexity — explained in 15 episodes

This episode discusses

The paper

Loop-Mamba: A Loop Mamba with Degradation-Aware and Shared Memory for Old Photo Restoration · Read on arXiv

China Academy of Information and Communications Technology

Old photographs often suffer from multiple coupled degradations, including scratches, cracks, fading, blur, noise, and missing regions, severely degrading both visual quality and semantic content. We propose Loop-Mamba, a lightweight loop-based state-space framework that formulates old photo restoration as progressive state evolution, where a persis- tent restoration state is continuously propagated and refined through iterative computation. Specifically, we introduce a Semantic-Guided Degradation Estimator (SGDE) to explicitly model heterogeneous degradations by jointly predicting local degradation maps and global degradation scores, providing degradation-aware guidance for state evolution. We further develop a Shared Structural Memory Mamba (S squared M- Mamba), which maintains a persistent restoration state across iterations, enabling persistent state evolution through shared structural memory for robust long-range structural reconstruction. Benefiting from first-order state recursion, Loop-Mamba propagates latent restoration states through recurrent tran- sitions instead of repeatedly stacking deep feature transformations, thereby alleviating gradient dilution while avoiding the computational overhead inherent in iterative CNN- and Transformer-based restoration frameworks. A lightweight multi-directional scanning strategy further enhances direc- tional information aggregation and preserves structural continuity. To better evaluate restoration quality, we introduce the task-oriented Old Photo Damage Recovery Score (ODRS), which jointly measures degradation recovery and structural reconstruction fidelity. Experimental results on the public SynOld benchmark demonstrate that Loop-Mamba consistently outperforms previous state-of-the-art methods across both conventional restoration metrics and the proposed ODRS.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Loop-Mamba: A Loop Mamba with Degradation-Aware and Shared Memory for Old Photo Restoration".

Jane: The gist The proposed Loop-Mamba framework formulates old photo restoration as a progressive loop state evolution process,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: This paper introduces the Loop-Mamba framework, which uses a loop-based state-space structure combined with degradation awareness and shared memory for old photo restoration.

Jane: It’s trying to move away from the old approach where you just map a degraded image directly to a clean one. They want this continuous process of state evolution instead.

Lu: The authors introduce the Semantic-Guided Degradation Estimator, or SGDE, which helps explicitly model those different kinds of damage by predicting local maps and global scores simultaneously.

Meng: So they’re not just looking at one type of noise or blur; they’re trying to understand the degradation in detail before they even start restoring anything.

Tom: Right. And that guidance then feeds into a loop state evolution process where the state updates based on that degradation awareness and some shared memory.

Jane: This shared memory part, which they call Shared Structural Memory Mamba or S2M-Mamba, is supposed to be crucial because it keeps accumulating structural representations across all the restoration iterations.

The paper's summary: Tom: The core of what Loop-Mamba does is formulating old photo restoration as a degradation-aware loop state evolution process where a persistent restoration state is continuously propagated and refined through iterative computation.

Jane: They start by saying that traditional methods treat it like a direct image-to-image mapping, but this new way focuses on the state evolution, which they call S t+one = F(S t, I d, D) where I d is the degraded image and D represents degradation <ref:2608.02346#pg1>.

Lu: The authors are motivated by observing that existing Mamba-based methods still rely on repeated feature refinement without explicitly modeling heterogeneous degradations or maintaining those persistent restoration states.

Meng: So they identified those two challenges—degradation modeling and state propagation—as the main problems they wanted to solve with this architecture.

Tom: And to fix them, they use the SGDE to get degradation-aware guidance for the state evolution, which is then propagated through a Mamba transition mechanism.

Jane: And on top of that, instead of starting from scratch every time, they have this Shared Structural Memory Mamba that continuously updates and propagates a shared structural state across restoration iterations.

The paper's improvements: Tom: They suggest several specific improvements to the methodology. First, they use a degradation-modulated state update formula where S et = S t + D l phi(S t) + Dg times psi(S t), which uses both local and global degradation information.

Jane: That's smart because it means the state update dynamically focuses on the severely degraded regions based on where the damage is happening right now.

Lu: Then they add a multi-directional state transition strategy, generating directional sequences and propagating them with the shared Mamba operator to enrich long-range interactions.

Meng: That sounds computationally intensive, but they claim it does this with negligible computational overhead, which is important for practical application on real hardware.

Tom: And finally, they have this joint reconstruction and degradation supervision objective, L = L rec + lambda L ddeg, where L rec handles the pixel fidelity and L ddeg supervises the predicted degradation scores.

Conclusion: Jane: So, to wrap things up, this Loop-Mamba framework is a lightweight state-space method for old photo restoration that uses progressive state evolution.

Tom: By combining semantic-guided degradation estimation with shared structural memory and direction-aware multi-scan propagation, they effectively model different types of damage while progressively refining the restoration states.

Lu: The authors conclude that Loop-Mamba achieves top scores across PSNR, SSIM, FSIM, and their task-oriented metric ODRS, while requiring substantially fewer FLOPs and MACs than other methods.

Meng: And they tested it on a Mi ten Lite Zoom smartphone using ONNX Runtime, which shows it’s feasible for real-world mobile deployment <ref:2608.02346#pg1>.

Tom: The progressive state evolution across five restoration cycles shows that the error maps become progressively smoother, which means residual errors are gradually eliminated over time.

Jane: It establishes a new way to think about restoring old photos by focusing on recurrent state evolution instead of just repeated feature reconstruction in this Loop-Mamba paper.

More episodes

← Home