Research map / Image generation / Image restoration
Multi-Focus Image Fusion via Explicit Defocus Blur Modelling
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 DMANet, an end-to-end multi-focus image fusion network that embeds an atom-based spatially-varying defocus blur model into a scale-recurrent defocus estimator, then fuses images with an uncertainty-aware fuser.
Problem
Deep learning multi-focus image fusion (MFIF) methods ignore the physical defocus blur model in network design, which limits interpretability and generalization.
Approach
Explicitly parameterize the defocus blur operator with spatially-invariant Gaussian PSF atoms and per-pixel coefficient maps, estimate these descriptors and initial focused images coarse-to-fine with a ConvLSTM scale-recurrent module, then fuse with an uncertainty-aware mask scheme that uses the initial focused images for ambiguous pixels.
Key results
- On MFI-WHU, DMANet reaches 42.898 dB PSNR and 0.9980 SSIM, versus 38.202 dB for the next-best DB-MFIF in the table, with the best ARank of 1.38.
- On Lytro, SIMIF, MFFW and OR-PAM, DMANet achieves the best ARank (1.2, 1.0, 1.4, 1.2 respectively) among compared methods.
- DMANet has 4.065M parameters and 0.03 s inference on 512x512 pairs, much smaller than GRFusion and FusionDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DMANet instead of DSIFT | MFIF · MFI-WHU | PSNR (dB): 39.396 -> 42.898 | Helped |
| Use DMANet instead of GRFusion | MFIF · Lytro | ARank: 2.0 -> 1.2 | Helped |
| Use DMANet instead of EAY-Net | MFIF · SIMIF | ARank: 2.6 -> 1.0 | Helped |
| Use DMANet instead of EAY-Net | MFIF · MFFW | ARank: 1.6 -> 1.4 | Helped |
| Use DMANet instead of GACN | MFIF · OR-PAM | ARank: 1.8 -> 1.2 | Helped |
| Use a coarse-to-fine multiscale DBE instead of one DBE scale (ablation) | DMANet · Lytro | ARank: 2 -> 1.4 | Helped |
| Add DBE (ablation) | DMANet · Lytro | ARank: 4.8 -> 1.4 | Helped |
| Add the scale-recurrent ConvLSTM structure (ablation) | DMANet · Lytro | ARank: 4.2 -> 1.4 | Helped |
| Add defocus estimation loss (ablation) | DMANet · Lytro | ARank: 2.6 -> 1.4 | Helped |
| Use uncertainty-aware UAF with γ=0.1 instead of binary masks with γ=0 (ablation) | DMANet · MFI-WHU | PSNR (dB): 42.781 -> 42.898 | Helped |
Surprising findings
- On MFFW, DMANet's QY (0.9578) is below EAY-Net's (0.9824), so it is not best on every metric.
- Reducing the UAF threshold gamma to 0 (binary masks) lowers PSNR only modestly (42.898 to 42.781 dB).
Limitations
- Gains on some metrics are marginal relative to strong baselines (e.g., Lytro NMI/MI close to GRFusion).
- Trained only on 10000 synthesized pairs, without pre-trained large-scale models.
- Evaluated only on S=2 source images in the main text; the general multi-source form is deferred to supplementary material.
Experimental setup
Datasets: DIV2K (Agustsson and Timofte 2017) Wang et al. 2019 synthesized pairs MFI-WHU Lytro SIMIF MFFW OR-PAM
Benchmarks: MFI-WHU Lytro SIMIF MFFW OR-PAM
Metrics: PSNR SSIM LPIPS NMI QG QM MI QY ARank
Baselines: DSIFT MFF MWGF GFF IFCNN GACN GRFusion FusionDiff DB-MFIF EAY-Net
Builds on: spatially-varying defocus deblurring model (Quan et al.) IFCNN EAY-Net
Closest papers by meaning
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.87)
- Blurry-Edges: Photon-Limited Depth Estimation from Defocused Boundaries (CVPR 2025 · similarity 0.87)
- DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal (CVPR 2025 · similarity 0.86)
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.86)
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.86)
- LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network (CVPR 2024 · similarity 0.85)
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.