Research map / Image generation / Image restoration
Hybrid-Supervised Dual-Search: Leveraging Automatic Learning for Loss-Free Multi-Exposure Image Fusion
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 HSDS-MEF, a bi-level differentiable search that jointly searches multi-exposure fusion network structures (WSRAS) and loss functions guided by a hybrid-supervised contrastive constraint using reference and natural images as positives.
Problem
Multi-exposure image fusion relies on manually designed network architectures and loss functions, and on pseudo reference images from SICE as ground truth, which caps performance and causes color distortion and exposure artifacts.
Approach
Jointly searches architecture and loss via bi-level optimization (DARTS-style), with a weighted structure refinement search that prunes low-weight sub-operations and retains top-P operations, and uses a hybrid contrastive constraint (simulated reference plus natural-light images as positives, source images as negatives) as the search criterion for the loss.
Key results
- On SICE, achieves VIF 1.585 vs. best baseline 1.433 (DeepFuse), and leads on SD, CC, TMQI, MS-SSIM and MEF-SSIM.
- Reported 10.61% VIF improvement in general scenario and 4.38% in no-reference scenario per abstract.
- Ablation: removing the hybrid contrastive constraint (w/o Γ_H) drops VIF from 1.585 to 1.308; DARTS baseline search yields VIF 1.429 vs. 1.585 for WSRAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hybrid-supervised contrast constraint instead of using no loss-function search (ablation) | HSDS-MEF · SICE | VIF: 1.308 -> 1.585 | Helped |
| Use only the reference contrast constraint (ablation) | HSDS-MEF · SICE | VIF: 1.585 -> 1.392 | Hurt |
| Use only the natural-image contrast constraint (ablation) | HSDS-MEF · SICE | VIF: 1.585 -> 1.176 | Hurt |
| Remove pruning refinement (ablation) | HSDS-MEF WSRAS · SICE | VIF: 1.585 -> 1.481 | Hurt |
| Remove weight retention (ablation) | HSDS-MEF WSRAS · SICE | VIF: 1.585 -> 1.426 | Hurt |
| Remove pruning refinement and weight retention (ablation) | HSDS-MEF WSRAS · SICE | VIF: 1.585 -> 1.429 | Hurt |
Surprising findings
- Only-reference contrastive constraint (Γ_R) gives second-best SICE metrics, and the natural-image constraint alone performs poorly quantitatively but helps visually on the no-reference set.
Limitations
- Relies on pseudo reference images from SICE for the reference-positive constraint.
- Search space restricted to lightweight convolution operations and a fixed set of loss candidates.
Experimental setup
Datasets: SICE MEF no-reference dataset COCO val2017
Benchmarks: SICE MEF dataset without reference
Metrics: SD VIF CC TMQI MS-SSIM MEF-SSIM EN Q^AB/F
Baselines: DeepFuse PMGI CF-NET AGAL TransMEF DPE-MEF HoLoCo DeFusion U2Fusion
Builds on: DARTS VGG-16 U2Fusion
Closest papers by meaning
- Conditional Controllable Image Fusion (NeurIPS 2024 · similarity 0.85)
- Region-Aware Exposure Consistency Network for Mixed Exposure Correction (AAAI 2024 · similarity 0.85)
- Ultra-High-Definition Dynamic Multi-Exposure Image Fusion via Infinite Pixel Learning (AAAI 2025 · similarity 0.85)
- MRFS: Mutually Reinforcing Image Fusion and Segmentation (CVPR 2024 · similarity 0.84)
- When Integral Meets Decomposition: A Signal-Level Self-Supervised Feature Decompose Paradigm for Multi-Modal Image Fusion (NeurIPS 2026 · similarity 0.84)
- Beyond Spatial-Domain Supervision: A Relation Constrained Space for Multi-Modal Image Fusion (NeurIPS 2026 · 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.