Does the original approach need exposure times?

It does not require them to reconstruct radiance, because it blends the input images directly rather than performing that reconstruction.

Must the images be aligned?

Corresponding content must be brought into agreement; otherwise blending can produce duplicated edges or other artifacts.

Is every fused image more realistic?

No. Weighting and blending choices can create unnatural appearance, so the result needs visual review.

AI Glossary / Computer vision and multimodal AI / Intermediate

Exposure fusion

Definition

Exposure fusion combines differently exposed images of a scene into a single display-oriented image by favoring useful content from each input. In the Mertens approach, local measures of contrast, color saturation, and well-exposedness guide multiscale blending. The method does not first reconstruct a physically calibrated high-dynamic-range radiance image.

What it means

One exposure may show good detail in a bright window while another better reveals furniture in a shadowed room. Fusion aims to bring useful parts of both into one picture that can be viewed directly.

This is a different objective from estimating scene radiance. The fused output is designed around visual qualities, and its pixel values should not automatically be interpreted as proportional to the amount of light in the original scene.

Exposure fusion is also distinct from exposure bracketing. Bracketing captures the input sequence; fusion combines it. Good capture and alignment remain important because a blending algorithm cannot assume that the same pixel location depicts the same object in every frame.

How it works

Start with aligned images. Compute local quality measures and combine them into weights for each image location. In the original Mertens formulation, adjustable exponents control the contribution of contrast, saturation, and well-exposedness within a product of measures.

Normalize the weights across the images so that they sum to one at a location. Rather than simply averaging original pixels using abruptly changing weights, build pyramids and blend information at multiple spatial scales. Reconstruct the final image from the combined levels.

The multiscale step helps manage visible transitions between inputs of different brightness. It does not make the method immune to ghosting, noise, or unsuitable weighting. Implementations can choose different defaults, so the active measures should be checked rather than inferred from the algorithm's name.

Three practical examples

1. Property photography: A photographer combines a stable exposure sequence of an interior to present both shaded room details and bright outdoor scenery.

2. Product presentation: A studio blends aligned exposures of a stationary object when different regions are rendered usefully in different frames, then checks the result against the intended appearance.

3. Computational imaging education: Students compare a direct weighted average with a pyramid-based blend to see how abrupt weight changes can create seams.

Think of it this way

Think of assembling a presentation from several photographs, each of which shows a different part of the scene clearly. Fusion chooses and blends contributions rather than trusting one photograph everywhere. The analogy is limited because the decisions are numerical and local; the method does not understand whether a chosen highlight or texture is important to the viewer.

Worked example

Consider a simplified location with normalized grayscale values of 0.1, 0.5, and 0.9 in three exposures. Suppose their unnormalized quality weights are 0.1, 0.8, and 0.1. They already sum to one, so a direct weighted average is 0.1 × 0.1 + 0.8 × 0.5 + 0.1 × 0.9 = 0.50.

At a different location, imagine values of 0.3, 0.8, and 1.0 with weights of 0.8, 0.19, and 0.01. The average is 0.24 + 0.152 + 0.01 = 0.402. The darkest exposure contributes most because the supplied weights favor it there.

These examples explain local weighting only; they are not a complete simulation of Mertens fusion. The actual multiscale process blends pyramid coefficients using corresponding weight-pyramid levels, so its reconstructed pixel need not equal the simple averages above.

Neither 0.50 nor 0.402 is an estimate of physical scene radiance. The values depend on the input rendering, quality measures, and blending. This distinction prevents a visually useful output from being mistaken for calibrated measurement data.

Illustrative case study

Fictional case study: A gallery photographs a display containing a dark frame and reflective glass. The imaging team captures several exposures with a fixed camera and fuses them for a catalog draft.

The draft reveals frame details, but a reflection changes between captures when someone moves nearby. The fused result contains an implausible doubled reflection. The team repeats the capture under controlled conditions and compares the new fusion with the best single exposure.

The catalog editor selects the version that represents the artwork and its presentation appropriately. The team retains the source sequence and processing settings. The case shows that useful local detail and overall visual credibility must both be reviewed; it does not assume fusion always wins the comparison.

Common misconceptions and mistakes

Exposure fusion does not necessarily create an HDR radiance map. A result that displays detail across a high-contrast scene can still be a conventional display-oriented image.

The well-exposedness measure is a heuristic preference for certain intensity values, not a universal judgment about correct artistic exposure. A deliberately dark subject may not benefit from being made brighter.

Contrast weighting can also favor noise or unwanted texture. More contrast is not always more useful information.

Pyramid blending does not automatically align moving objects. Registration and handling of scene changes are separate concerns, and difficult input sequences may require recapture or manual intervention.

Business use

Fusion is useful when a team wants a practical, visually balanced image without needing a calibrated radiance representation. It can support photographic editing and serve as a component or baseline in computational imaging workflows.

Evaluate seams, ghosting, color shifts, noise, and consistency across a set of images. A good workflow also records capture conditions and protects originals. If downstream software relies on brightness as a physical measurement, evaluate a calibrated imaging approach instead of assuming a pleasing fused image preserves that relationship.

Check your understanding

Two quality weights are 2 and 6. What are their normalized values?

Answer: 0.25 and 0.75, obtained by dividing each by their sum, 8. These weights describe contribution, not confidence in physical radiance.

Related terms

Exposure bracketing
Tone mapping
Image registration
HDR radiance map
Laplacian pyramid
multiscale blending
ghosting

Learn next

Learn normalized weighted averaging, then examine why weight changes cause seams across differently exposed images. Compare direct blending with pyramid blending on a static bracket and inspect the behavior around strong boundaries.

Sources

Mertens, Kautz, and Van Reeth, Exposure Fusion, author-hosted paper: https://www.tommertens.com/old-academic/papers/exposure_fusion_reduced.pdf

OpenCV, MergeMertens official reference: https://docs.opencv.org/4.13.0/d7/dd6/classcv_1_1MergeMertens.html

Both primary sources were opened. Numerical weighting examples and the gallery scenario are original illustrations.

Previous: Exposure bracketing Next: Expressiveness