Research map / Label-efficient and robust learning / Recognition and long-tail
Region-Aware Exposure Consistency Network for Mixed 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 RECNet, a region-aware network for mixed-exposure image correction that uses a region-aware de-exposure module, a mixed-scale restoration unit, and an exposure contrastive regularization loss.
Problem
Correcting images with mixed exposure, where a single image contains both overexposed and underexposed regions, which existing global unified exposure correction methods handle poorly.
Approach
Segment over- and underexposed regions with a learned exposure mask, apply mask-aware instance normalization to map regional features into an exposure-invariant space, restore lost detail with multi-scale depth-wise convolution and channel-wise self-attention, and regularize with a VGG-feature contrastive loss enforcing intra-region consistency and inter-region exposure continuity.
Key results
- RECNet reaches 23.69 dB PSNR / 0.8699 SSIM average on MSEC, outperforming ECLNet+ERL (22.70 / 0.8655).
- On LCDP mixed-exposure data, RECNet achieves 23.54 dB PSNR / 0.8468 SSIM, a marginal gain over LCDPNet (23.24 / 0.8420).
- Ablation shows the full model improves PSNR from 22.09 (IN baseline) to 23.54 dB on LCDP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MaIN instead of IN (ablation) | RECNet · LCDP | PSNR: 22.09 -> 22.41 | Helped |
| use MaIN instead of IN (ablation) | RECNet · LCDP | SSIM: 0.8292 -> 0.8305 | Helped |
| add multi-scale convolution (MSC) to MaIN (ablation) | RECNet · LCDP | PSNR: 22.41 -> 22.98 | Helped |
| add multi-scale convolution (MSC) to MaIN (ablation) | RECNet · LCDP | SSIM: 0.8305 -> 0.8386 | Helped |
| add channel-wise self-attention (CSA) to MaIN (ablation) | RECNet · LCDP | PSNR: 22.41 -> 23.14 | Helped |
| add channel-wise self-attention (CSA) to MaIN (ablation) | RECNet · LCDP | SSIM: 0.8305 -> 0.8394 | Helped |
| add both MSC and CSA to MaIN (ablation) | RECNet · LCDP | PSNR: 22.41 -> 23.37 | Helped |
| add both MSC and CSA to MaIN (ablation) | RECNet · LCDP | SSIM: 0.8305 -> 0.8431 | Helped |
| add exposure contrastive regularization (ECR) (ablation) | RECNet with MaIN, MSC, and CSA · LCDP | PSNR: 23.37 -> 23.54 | Helped |
| add exposure contrastive regularization (ECR) (ablation) | RECNet with MaIN, MSC, and CSA · LCDP | SSIM: 0.8431 -> 0.8468 | Helped |
| use RECNet instead of FECNet | exposure correction methods · MSEC Average | PSNR: 23.12 -> 23.69 | Helped |
| use RECNet instead of MSEC+DA | exposure correction methods · SICE Average | SSIM: 0.7093 -> 0.7309 | Helped |
Surprising findings
- On the SICE underexposure split, ECLNet+ERL-style and MSEC+DA methods achieve higher SSIM than RECNet (MSEC+DA 0.7546 vs RECNet 0.7254), so RECNet is not best on every metric.
Limitations
- Gains over LCDPNet on the LCDP mixed-exposure dataset are marginal.
- Mixed-exposure evaluation relies on a single synthetic dataset rendered from MIT-Adobe FiveK.
Experimental setup
Datasets: MSEC SICE LCDP
Benchmarks: MSEC SICE LCDP
Metrics: PSNR SSIM
Baselines: RetinexNet RUAS SCI LCDPNet FECNet DRBN-ENC MSEC MSEC+DA ECLNet ECLNet+ERL
Builds on: LCDPNet ECLNet DRBN Retinex theory Exposure contrastive regularization
Closest papers by meaning
- Exposure-slot: Exposure-centric Representations Learning with Slot-in-Slot Attention for Region-aware Exposure Correction (CVPR 2025 · similarity 0.87)
- CLIP-RestoreX: Restore Image Structure and Perception in Exposure Correction (AAAI 2025 · similarity 0.85)
- Hybrid-Supervised Dual-Search: Leveraging Automatic Learning for Loss-Free Multi-Exposure Image Fusion (AAAI 2024 · similarity 0.85)
- Color Shift Estimation-and-Correction for Image Enhancement (CVPR 2024 · similarity 0.84)
- Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling (CVPR 2024 · similarity 0.84)
- InstructRestore: Region-Customized Image Restoration with Human Instructions (NeurIPS 2025 · similarity 0.84)
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.