Research map / Architectures and efficiency / Attention and transformers
ECMamba: Consolidating Selective State Space Model with Retinex Guidance for Efficient Multiple Exposure Correction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ECMamba, a dual-branch Retinex-guided Mamba network with a feature-aware 2D selective scan (Retinex-SS2D) for multi-exposure and under-exposure correction.
Problem
Exposure correction for both over- and under-exposed images is hard for existing deep models, which rarely integrate Retinex theory fully and often trade off performance against efficiency (especially transformers).
Approach
A Retinex estimator maps the input into intermediate reflectance-like and illumination-like spaces that are restored by two separate Mamba-based branches; within the Mamba block, a feature-aware 2D scan orders tokens by deformable-convolution activation frequency rather than by fixed direction, so that important features come first in the S6 sequence.
Key results
- On the ME dataset, ECMamba reaches 23.76 average PSNR and 0.879 SSIM, beating the second-best LACT (23.57 / 0.869).
- On SICE, ECMamba reaches 22.05 average PSNR and 0.736 SSIM, versus 21.22 / 0.685 for the second-best method.
- On LOLv2-real, ECMamba reaches 29.24 PSNR vs 28.19 for LLFlow-SKF, using about 1.75M parameters (about 4.4% of LLFlow-SKF's 39.91M).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the illumination-restoration branch M_L (ablation) | ECMamba · SICE average | PSNR: 21.55 -> 22.05 | Helped |
| add the illumination-restoration branch M_L (ablation) | ECMamba · SICE average | SSIM: 0.721 -> 0.736 | Helped |
| add the illumination-restoration branch M_L rather than widening M_R (ablation) | ECMamba · SICE average | PSNR: 21.63 -> 22.05 | Helped |
| add the Retinex estimator E (ablation) | ECMamba · SICE average | PSNR: 21.12 -> 22.05 | Helped |
| add the Retinex estimator E (ablation) | ECMamba · SICE average | SSIM: 0.695 -> 0.736 | Helped |
| use the ECMamba module rather than ViT (ablation) | ECMamba · SICE average | PSNR: 21.88 -> 22.05 | Helped |
| use the ECMamba module rather than Retiformer (ablation) | ECMamba · SICE average | PSNR: 21.35 -> 22.05 | Helped |
| use FA-SS2D rather than Cross-Scan (ablation) | ECMamba · SICE average | PSNR: 21.69 -> 22.05 | Helped |
| use FA-SS2D rather than Cross-Scan (ablation) | ECMamba · SICE average | SSIM: 0.716 -> 0.736 | Helped |
| use ECMamba instead of LACT | multi-exposure correction · ME average | PSNR: 23.57 -> 23.76 | Helped |
| use ECMamba instead of LLFlow-SKF | under-exposure correction · LOLv1 | PSNR: 26.80 -> 27.69 | Helped |
| use ECMamba instead of LLFlow-SKF | under-exposure correction · LOLv2-real | PSNR: 28.19 -> 29.24 | Helped |
Surprising findings
- Removing the Retinex estimator and directly restoring the image lowers performance notably (21.12 vs 22.05 average PSNR on SICE), indicating the intermediate space matters.
Limitations
- Struggles with extreme exposure cases (extremely dark or over-exposed) due to information loss.
- Ablations are reported only on the SICE dataset.
Experimental setup
Datasets: ME SICE LOLv1 LOLv2-real LOLv2-synthetic
Benchmarks: ME SICE LOLv1 LOLv2-real LOLv2-synthetic
Metrics: PSNR SSIM Param (M)
Baselines: ZeroDCE RUAS URetinexNet KinD LLFlow LLFlow-SKF DRBN FECNet Retiformer LACT
Builds on: Mamba VMamba cross-scan Retiformer LLFlow FECNet
Closest papers by meaning
- MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space (NeurIPS 2024 · similarity 0.86)
- EfficientVMamba: Atrous Selective Scan for Light Weight Visual Mamba (AAAI 2025 · similarity 0.84)
- IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective (AAAI 2025 · similarity 0.83)
- Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model (ICLR 2025 · similarity 0.83)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.83)
- OSMamba: Omnidirectional Spectral Mamba with Dual-Domain Prior Generator for Exposure Correction (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.