Is deghosting the same as deblurring?

No. It addresses inconsistent contributions across images; deblurring estimates detail affected by blur.

Why can a corrected region become noisy?

The method may rely on fewer usable observations there, reducing the benefit of combining frames.

Does every method choose one reference image?

No. Reference selection is one strategy among several alignment and fusion approaches.

AI Glossary / Computer vision and multimodal AI / Advanced

Deghosting

Definition

Deghosting reduces duplicate, translucent or inconsistent structures created when multiple images containing motion or misalignment are combined. In HDR and exposure fusion, it addresses disagreement between exposures captured at different moments. In image stitching, related techniques choose or reconcile content in overlaps so moving objects do not appear as incompatible fragments or repeated copies.

What it means

Combining photographs assumes that their contributions can describe a coherent result. If a person moves between captures, an ordinary average may place a faint person at each recorded position. Deghosting tries to avoid that inconsistency. It may choose one observation, reduce unreliable contributions or align corresponding content before fusion. These choices can sacrifice some information from the other inputs. The result is a constructed composite, and a clean appearance does not prove that every part represents one simultaneously captured moment.

How it works

Methods compare images or features, identify inconsistent regions and determine which observations should contribute. Strategies include reference-based rejection, motion alignment, seam selection and learned fusion weights. Exposure changes complicate direct comparisons because brightness differences need not indicate movement. Occlusion and saturation make the task harder by removing reliable correspondences or useful measurements. The precise method depends on whether the goal is HDR reconstruction, exposure fusion or stitching. Deghosting is separate from tone mapping and is not simply another name for reducing blur in one photograph.

Three practical examples

A photographer merges an exposure bracket containing moving leaves and checks whether the final image has doubled edges or unnatural patches around them.
A panorama editor chooses an overlap seam that avoids splitting a walking person across images captured at different times.
An imaging team evaluates a fusion method on controlled moving objects, checking both visual consistency and whether useful exposure information is lost.

Think of it this way

Imagine combining attendance photographs taken a few seconds apart. One person has moved from the left side to the right. Averaging both pictures can create two faint versions, while selecting one photograph for that person gives a single appearance. The choice solves one visual inconsistency but determines which moment is represented. Deghosting makes comparable decisions through masks, alignments or fusion rules; it cannot assume all source observations describe an unchanged scene.

Worked example

In an original binary-row example, frame A is [0,1,0,0] and frame B is [0,0,1,0], representing a bright object that moved one position. An equal blend gives [0,0.5,0.5,0], producing two partial appearances. Selecting frame A for the affected region gives [0,1,0,0]. Alternatively, if a valid motion estimate aligns B's object with A's position before blending, the aligned result can retain one object. These idealized choices assume known motion and uncomplicated visibility. Real exposure brackets may contain saturation, changing backgrounds or occlusions, so selecting or aligning content can lose information or introduce other artifacts.

Illustrative case study

A fictional travel-media team combines several exposures of a square with pedestrians. A strong deghosting setting removes obvious duplicates but leaves a noisy patch where only one usable exposure contributes. The team compares reference choices and reviews people, shadows and borders at the intended output size. It retains the source bracket and describes the result as a composite when the capture context matters. This scenario illustrates the tradeoff between consistency and the benefits of combining multiple observations. It does not claim that stronger deghosting always improves the image or that the processed scene is a precise record of one instant.

Common misconceptions and mistakes

Deghosting does not guarantee removal of every motion artifact. A region selected from one exposure may have less dynamic-range coverage or more noise than a successfully fused static region. Another mistake is treating all intensity disagreement as movement; exposure and illumination also change values. A reference frame with clipped detail may provide poor guidance. Inspect moving-object boundaries and partially occluded regions, not just the center of large objects. Finally, display-panel ghosting and optical reflections are different phenomena, even though the same everyday word is used.

Business use

The concept is relevant to HDR photography, panoramas and other multi-image workflows. A business should decide whether the output is intended for attractive presentation, documentation or quantitative analysis, since each requires different validation. Review both duplicated structures and less obvious losses of detail around corrected regions. Preserve source images and relevant settings when later investigation matters. Automated acceptance based only on global image quality can overlook small but important moving objects. Deghosting should be assessed as part of the complete capture and fusion process, including alignment and exposure coverage.

Check your understanding

Why did the simple average show two partial objects?

Answer: The bright object occupied different positions in the two inputs.

What did selecting frame A accomplish?

Answer: It retained one temporal observation in that region.

What information might that choice sacrifice?

Answer: Useful detail or exposure coverage available only in other frames.

Related terms

Exposure fusion
Image registration
Motion estimation
HDR imaging
occlusion
seam selection

Learn next

Blend two photographs of a moving object and locate the resulting duplicates. Compare selecting one source region with aligning the object before blending. Then inspect how occlusion and exposure clipping limit both approaches.

Sources

Hu and colleagues, HDR Deghosting: How to Deal with Saturation?: https://www.cv-foundation.org/openaccess/content_cvpr_2013/papers/Hu_HDR_Deghosting_How_2013_CVPR_paper.pdf Patch-Based Correlation for Deghosting in Exposure Fusion, primary publication abstract: https://www.sciencedirect.com/science/article/abs/pii/S0020025516319053 The binary rows, attendance-photo analogy and travel-media case are original educational illustrations. Deghosting methods differ in how they resolve conflicts among inputs.

Previous: Degenerate dimension Next: Degree adverb