2013 paper record · 2026 reference reconstruction

Blurred + noisy restored

A framework-free gallery for the preserved paper archive and a disparity-guided regional restoration reference implementation. Each comparison is ordered blurred, noisy auxiliary, restored, then optional reference.

Click an image, or focus it and press Enter, to enlarge it.

What this is

Published journal article · August 2013 Disparity-based space-variant image deblurring Signal Processing: Image Communication · 28(7) · 792–808 doi:10.1016/j.image.2013.04.005

The preserved paper PDF and historical 2009–2011 README figures are research records. The 2026 Python package is a reproducible modern reference reconstruction, not the exact deleted original implementation and not bit-identical to historical outputs.

2013 paper record · 2026 reference reconstruction

2013 paper record

  1. Blurred/noisy capture pair
  2. Harris/LK disparity
  3. Color graph-cut over-segmentation
  4. Adjacent disparity-region merge
  5. Regional gradient-domain Tikhonov PSF
  6. Kurtosis validation and replacement
  7. L1/FFT regional restoration and merge

2026 reference reconstruction

  1. Deterministic RANSAC homography registration
  2. Graph-cut and disparity-median region merge
  3. Regional Tikhonov kernels and TV/L1 FFT deconvolution
  4. Normalized feather blending
  5. Aspect-ratio-safe portrait, landscape, and odd-size processing
  6. Optional C++17 and CUDA/cuFFT deconvolution backends
  7. Optional guided noisy-detail fusion

Guided noisy-detail fusion is a 2026 enhancement, not a claim about the deleted 2013 code. This reconstruction is not bit-identical to historical outputs.

Living-room restoration walkthrough

Mat3 is an authorized Hyeon Sang Jeon living-room photograph rendered with the selected HPO preset. It has no clean ground truth: the 0.874256 proxy score and 0.919059 proxy SSIM use the registered, denoised noisy auxiliary, not PSNR.

Selected HPO: 12 color labels, disparity threshold 1.0, 17px kernels, 192px patches, Tikhonov 0.001, data weight 60, beta 8, feather sigma 12, unsharp amount 0.3, and guided detail-fusion amount 0.5.

Mat3 living-room blurred input photograph.
1. Blurred input
Mat3 short-exposure noisy auxiliary photograph.
2. Noisy auxiliary
Noisy auxiliary after RANSAC homography registration to the blurred image.
3. Registered noisy image
Harris feature and disparity-vector visualization over the Mat3 scene.
4. Feature/disparity vectors
Initial color graph-cut over-segmentation regions.
5. Initial graph-cut regions
Regions merged by adjacent disparity medians.
6. Disparity-merged regions
Regional point-spread-function kernel visualization.
7. Regional PSF kernels
Regional TV L1 FFT deconvolution intermediate.
8. Deconvolved intermediate
Final HPO-optimized Mat3 living-room restoration.
9. Final HPO restored result
Compact Mat3 comparison ordered blurred, noisy auxiliary, and final restored image.
Blurred + noisy → restored
Historical object-motion photograph comparison: blurred, explicitly noise-amplified auxiliary, and restored images.

Historical photograph: object motion

Metric
No-reference proxy
Objective (full-resolution)
0.632649

The middle panel adds deterministic σ=0.02 Gaussian noise for visible separation; the original historical auxiliary remains archived.

Historical low-light building photograph comparison: blurred, noisy, and restored images.

Historical photograph: low-light building

Metric
No-reference proxy
Objective (full-resolution)
0.640904
Historical parking photograph comparison: blurred, noisy, and restored images.

Historical photograph: parking

Metric
No-reference proxy
Objective (full-resolution)
0.707996

Ringing-aware result: 35% regional restoration plus up to 65% confidence-guided, denoised noisy-image structure.

Derived GoPro Flower comparison: blurred, noisy, restored, and reference images.

External GoPro-derived benchmark

PSNR
25.3705
SSIM
0.8816
Objective (full-resolution)
0.694603

Attribution and license

Metrics and selected HPO configurations are from the committed public benchmark manifest output. The historical-photograph values are conservative no-reference proxy objectives, not PSNR or SSIM.

Reproduce

uv sync
uv run disparity-deblur-benchmark \
  --manifest benchmarks/manifests/public.json \
  --dataset-root benchmarks/public-assets \
  --output-dir output/showcase

NumPy remains the canonical backend. See the optional C++ and CUDA acceleration manual for explicit native builds, cuFFT usage, limitations, and reproducible timing.