Abstract
Exposure Fusion (EF) is a widely used technique for blending bracketed image sequences into a single high-quality image. However, the standard Mertens algorithm often suffers from out-of-range clipping and low-frequency halos. This paper presents Extended Exposure Fusion (EEF), which utilizes a remapping function to restrain dynamic range and deep Laplacian pyramids to eliminate artifacts. Furthermore, the authors introduce Simulated Exposure Fusion (SEF), extending this framework to single-image contrast enhancement by generating virtual exposure sequences.
Introduction
Exposure Fusion (EF) simplifies the high-dynamic-range (HDR) imaging pipeline by directly blending multiple exposures into an 8-bit displayable image, bypassing the need for intermediate 32-bit HDR files. Despite its popularity, the original EF algorithm is prone to two major issues: out-of-range artifacts and low-frequency halos. Out-of-range artifacts occur when fused pixel values exceed the [0, 1] range; clipping these values destroys detail, while global normalization reduces overall contrast.
Methodology
The proposed Extended Exposure Fusion (EEF) addresses these flaws through two key innovations:
- Restrained Dynamic Range: Each input image is processed through a remapping function $g(t, k)$ that "squeezes" its dynamic range. This provides sufficient headroom during the fusion process, ensuring the final result stays within the displayable [0, 1] range without requiring destructive clipping or global rescaling.
- Deep Laplacian Pyramids: By resolving the out-of-range issue, the authors can employ much deeper pyramids (down to a 1x1 residual). This effectively eliminates the low-frequency halos that plague shallower pyramids.
Simulated Exposure Fusion (SEF)
To enhance single images, the authors propose Simulated Exposure Fusion (SEF). This method analyzes the image histogram to determine a median value and generates a "virtual" bracketed sequence using contrast amplification factors. The sequence is then fused using the EEF algorithm in the HSV colorspace, specifically targeting the Value channel to preserve color fidelity.
Key Insights
A central finding is the Contrast Paradox: reducing the dynamic range of input images before fusion actually yields higher local contrast in the output by preventing clipping. Additionally, the study emphasizes that halo artifacts are a byproduct of pyramid depth; pushing the decomposition to its limit effectively removes these artifacts.
Results and Discussion
Qualitative evaluations show that EEF successfully recovers details in overexposed regions and eliminates halos in high-contrast scenes. Quantitatively, SEF consistently achieves high scores on the Tone-Mapping Quality Index (TMQI), outperforming traditional methods like CLAHE and MSR.
While SEF is highly effective at enhancing detail, it has limitations regarding sensor noise. Because it is a contrast enhancement technique, it amplifies existing noise in extremely dark regions at the same rate as detail. Future work could involve integrating a denoising step into the virtual exposure generation or the fusion process to make the algorithm more robust for low-light photography.
Conclusion
EEF and SEF provide a robust framework for both multi-exposure fusion and single-image contrast enhancement. By addressing the fundamental limitations of the Mertens algorithm—specifically clipping and halos—this work enables high-quality image generation with improved local contrast and superior artifact suppression.
Summary updated