Improved Image Selection for Stack-Based HDR Imaging

• Existing fixed-bracketing methods require the photographer to specify N and may either acquire unnecessary images or miss irradiance in challenging scenes.Source: Introduction · p. 1
• Adaptive methods incorporate scene irradiance and camera noise, but their computational costs range from polynomial to NP-complete or empirically high.Source: Introduction · p. 1
• The proposed method is adaptive to irradiance, models camera noise, and formulates exposure selection as a polynomially solvable set covering problem.Source: Introduction · p. 1
• Across 110 benchmark scenes, the method improves HDR images over state-of-the-art methods using mean squared error, a pixel-based metric, and two perception-based metrics.Source: Introduction · p. 1

• RAW images are linearly related to scene radiance and typically have higher dynamic range than nonlinear 8-bit JPEG images.Source: Background · p. 1
• A radiometric response function maps scene radiance to JPEG pixel values, while its inverse estimates RAW values from JPEG observations.Source: Background · pp. 1–2
• The noise function estimates signal-to-noise ratio from RAW values using SNR(dB) = 20·log10(µ/σ), with exposure and ISO affecting SNR monotonically before saturation.Source: Background · p. 2
• The HDR histogram is estimated by sweeping shutter speeds in low-resolution live-preview JPEGs and combining their LDR histograms in the cumulative distribution function domain.Source: Background · p. 2

Selection as Set Covering

• Each low-resolution JPEG is swept across shutter speeds, providing preview images that can be acquired without extra cost when estimating scene range or an HDR histogram.Source: Selection as Set Covering · p. 2
• A pixel is accurately captured when its grayscale value lies in [Imin,Imax], defined using an SNR threshold and avoidance of multi-channel saturation.Source: Selection as Set Covering · p. 2
• The set-covering matrix represents pixels as rows, exposure settings as columns, and entries indicating whether each exposure accurately captures each pixel.Source: Selection as Set Covering · p. 2
• Unit costs minimize image count, whereas costs tj + tover minimize total capture time including acquisition overhead.Source: Selection as Set Covering · p. 2
• Although general set covering is NP-complete, the exposure instance has the consecutive-ones property, enabling polynomial-time selection and reduction rules.Source: Selection as Set Covering · pp. 2–3
• The selected exposures are acquired as high-resolution RAW images and combined into one HDR image using existing techniques.Source: Selection as Set Covering · p. 3
• A fixed higher ISO can be supported when minimum shutter speeds are needed for camera-shake reduction or dynamic scenes, with ISO incorporated throughout the pipeline.Source: Selection as Set Covering · p. 3

Evaluation design

• The evaluation is extensive, comparing four representative state-of-the-art methods across 110 benchmarks and multiple performance measures against ground-truth HDR images.Ground truth was constructed from the available high-resolution RAW images, excluding images that would only add noise.Source: Experimental Evaluation · p. 3

HDR Photographic Survey

• The set covering method improves all three performance measures on 105 HDR Photographic Survey benchmarks despite only nine candidate images per scene.The measures are quality correlate, visual difference prediction, and mean squared error.Source: Experimental Evaluation · p. 3

Canon benchmarks

• On five Canon benchmarks with 55 candidate exposures, the set covering method consistently achieves excellent results and captures challenging scenes that existing methods fail to fully capture.The candidates are spaced one-third of a stop apart, providing a more extensive exposure range.Source: Experimental Evaluation · p. 4

Speed

• The method is fast: MATLAB implementation required approximately 2/3 second for pixel classification and 1/3 second for image-set selection, versus 82.4 seconds for Hasinoff et al..Fast methods threshold acceptable SNR for each LDR image, enabling greedy selection; slow methods threshold the final HDR image, increasing computational complexity.Source: Experimental Evaluation · p. 4

Number of selected images

• Across all 110 benchmarks, the proposed set covering method selected 3 images at the median and 4 at the 75th percentile; its accuracy improvement was not at the expense of increasing capture time for most scenes.Barakat et al. selected 4 and 4 images, while Hasinoff et al. selected 3 and 3.Source: Experimental Evaluation · p. 4

Conclusion

• The technique minimizes the set of images to combine while ensuring that the resulting HDR image faithfully captures scene irradiance.This addresses the acquisition and processing costs of stack-based HDR imaging.Source: Conclusion · p. 4
• For two dark-region scenes, Hasinoff et al.’s method selected 76 multiple captures at the longest shutter speed, so the exposure sets were restricted to single captures.The authors note that these displayed results may be somewhat worse than they otherwise would have been.Source: Conclusion · p. 4
• Figure 3 evaluates each selection method across HDR Photographic Survey benchmarks using quality correlate, visually significant differences, and mean squared error.Within each method, each performance measure is sorted from smallest to largest.Source: Figure 3 · p. 4, Conclusion · p. 4
• The proposed set covering method has the highest quality and lowest error in the illustrated benchmark comparisons.Figure 4 compares visual differences, quality correlate, mean squared error, and selected-image counts across exposure-selection methods.Source: Conclusion · p. 4, Figure 4 · p. 5
• Across 110 benchmark scenes, the method improves HDR images against ground truth using pixel-based and perception-based metrics.The reported metrics include mean squared error, a pixel-based measure, and two perception-based measures.Source: Conclusion · p. 4

Author Biography

• Peter van Beek has research interests spanning AI, optimization, constraint programming, and applied machine learning.He has been a faculty member in Computer Science at the University of Waterloo since 2000.Source: Author Biography · p. 6