Research map / Image generation / Image restoration
Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model
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.
Introduces the QPDD dataset of 4,935 quad-pixel defocus/all-in-focus RAW/sRGB pairs and LMNet, a two-branch Mamba-based encoder-decoder with a Local-gate assisted Mamba Block for defocus deblurring on QP and DP data.
Problem
Defocus deblurring with spatially varying blur has been studied mainly with dual-pixel data; quad-pixel (QP) based defocus deblurring is scarce due to lack of datasets, and Mamba-based models suffer from local pixel forgetting and channel redundancy.
Approach
Capture QP defocus/all-in-focus pairs by sweeping the VCM lens and aligning with ECC; design LMNet with a two-branch encoder (one branch per sub-aperture pair), a Simple Fusion Module, and LAMB, which parallelly combines a VSSM (global) and a Local Gate Module (CA+SA gating) with learnable scales.
Key results
- On DPDD, LMNet reaches 27.25 dB combined PSNR with 18.18M params, matching LDP (27.24 dB) and beating K3DN (27.06 dB) and Restormer (26.66 dB).
- On QPDD, four-view input (OursQ) achieves 36.75 dB PSNR (RAW) versus 36.48 dB for two-view input, outperforming Restormer and LaKDNet two-view sRGB retrained baselines (34.36 and 34.64 dB).
- Ablation on DPDD: two-branch encoder adds 0.29 dB PSNR over single encoder; SFM adds 0.15 dB over simple addition.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use four-view L/R and T/B input instead of two-view L/R input | LMNet on QPDD · QPDD test, RAW | PSNR: 36.48 -> 36.75 | Helped |
| Use four-view L/R and T/B input instead of two-view L/R input | LMNet on QPDD · QPDD test, sRGB | PSNR: 36.25 -> 36.64 | Helped |
| Use LMNet instead of Restormer | DPDD defocus deblurring · DPDD test, combined | PSNR: +0.59 dB | Helped |
| Use a two-branch encoder instead of a single encoder (ablation) | LMNet · DPDD test | PSNR: 26.96 -> 27.25 | Helped |
| Add SFM instead of using element-wise addition for branch fusion (ablation) | LMNet · DPDD test | PSNR: +0.15 dB | Helped |
| Add LGM to VSSM (ablation) | LAMB · DPDD test | PSNR: 26.77 -> 27.25 | Helped |
| Add VSSM to CAB (ablation) | LAMB · DPDD test | PSNR: 26.76 -> 26.95 | Helped |
| Replace CAB with LGM alongside VSSM (ablation) | LAMB · DPDD test | PSNR: 26.95 -> 27.25 | Helped |
| Use LGM instead of CAB without VSSM (ablation) | LAMB · DPDD test | PSNR: 26.76 -> 26.79 | Helped |
Surprising findings
- Four-view QP input improves PSNR by only 0.27 dB (RAW) over two-view input within LMNet, a modest gain.
- LaKDNet trained on DPDD generalizes with the same PSNR as Restormer (29.53 dB) on QPDD while LMNet reaches 30.37 dB.
Limitations
- QPDD consists mainly of planar indoor scenes with limited depth-of-field diversity.
- Model is not designed for mobile or embedded deployment; authors plan a lighter network as future work.
- Dataset captured with a fixed-aperture QP sensor using VCM focus sweeps and ECC alignment, which may introduce residual misalignment.
Experimental setup
Datasets: DPDD QPDD
Benchmarks: DPDD QPDD
Metrics: PSNR SSIM MAE Params (M) FLOPs (G)
Baselines: DPDNet RDPD UFormer IFAN MIRNetV2 DRBNet BaMBNet Restormer LaKDNet K3DN LDP
Builds on: Mamba VMamba Restormer MambaIR DPDNet
Closest papers by meaning
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.89)
- LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network (CVPR 2024 · similarity 0.88)
- Multi-Focus Image Fusion via Explicit Defocus Blur Modelling (AAAI 2025 · similarity 0.87)
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.87)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.85)
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (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.