Research map / Image generation / Image restoration
Focus Stacking with High Fidelity and Superior Visual Effects
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 a two-stage foreground segmentation and diffusion elimination architecture, modeling the optical PSF-based diffusion at the transition region to produce all-in-focus images with both high fidelity and good visual effects.
Problem
Focus stacking must yield all-in-focus images that are both high-fidelity to source images and free of visual defects, but decision-based methods leave halo/transition-region artifacts and reconstruction methods cause color bias.
Approach
Models the diffusion component at the transition region as the smoothed color difference between foreground and background convolved with a PSF; a foreground segmentation branch produces a decision map, and a kernel estimation branch estimates the PSF, then the diffusion component is subtracted from the initial synthesis.
Key results
- On the simulated dataset, the method reports QMI 1.238, QAB/F 0.906, QY 0.898, mIOU 0.966 and PSNR 36.9, the best among compared methods.
- Diffusion elimination removes the transition-region slope seen in initial syntheses, and an SSIM of 0.9571 and PSNR of 35.46 are reported for kernel estimation consistency on DIV2K.
- Applying the diffusion elimination to DRPL's initial synthesis visibly reduces the transition band artifact, showing the module generalizes to other decision-based methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add diffusion elimination to the initial synthesis (ablation) | Ours · Focus stacking transition regions | visual effects | Helped |
| apply diffusion elimination to DRPL's initial synthesis (ablation) | DRPL · Focus stacking transition regions | visual effects | Helped |
| use the proposed method instead of Liu et al. | Focus stacking · Simulated dataset test | QAB/F: 0.904 -> 0.906 | Helped |
| use the proposed method instead of Liu et al. | Focus stacking · Simulated dataset test | PSNR: 36.3 -> 36.9 | Helped |
Surprising findings
- Decision-based methods ignore transition-region artifacts because arbitrary pixel selection there barely changes fidelity metrics.
Limitations
- Relies on accurate foreground masks; inaccurate segmentation of monotonous regions can produce incorrect diffusion estimates.
- ZMFF results omitted on the simulated dataset due to its long inference time.
- Real-world evaluation lacks ground truth, so transition-region quality is judged qualitatively and by QCV.
Experimental setup
Datasets: AM2K DIV2K LytroDataset BetaFusion
Benchmarks: simulated focus stacking test set (1.8k) BetaFusion LytroDataset
Metrics: QMI QAB/F QY mIOU PSNR QCV
Baselines: DeFusion Liu et al. MFFNet ZMFF DRPL GEU-Net DTMNet Wang et al.
Builds on: ResNet U-Net Conditional random field
Closest papers by meaning
- MACRO: Training-free Multi-plane Attention for Closeup Render Optimization (NeurIPS 2026 · similarity 0.81)
- DualFocus: Depth from Focus with Spatio-Focal Dual Variational Constraints (NeurIPS 2025 · similarity 0.81)
- Era3D: High-Resolution Multiview Diffusion using Efficient Row-wise Attention (NeurIPS 2024 · similarity 0.81)
- Scene Splatter: Momentum 3D Scene Generation from Single Image with Video Diffusion Model (CVPR 2025 · similarity 0.81)
- UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion (CVPR 2025 · similarity 0.80)
- Generative Image Layer Decomposition with Visual Effects (CVPR 2025 · similarity 0.80)
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.