Why use more than one transition width?
When two registered photographs overlap, a hard switch can expose a brightness difference at the seam. A wide cross-fade distributes that difference, but can also superimpose two displaced edges over a wide region. These are different spatial scales of disagreement. A multiscale blend gives the broad variations a broad transition and gives fine details a narrower transition. It does not estimate alignment, exposure, or the physically correct scene; it combines images under those choices.
We use a deliberately small finite example so every output can be checked. Pixel values represent a single linear-light channel. Apply the construction independently to channels, using the same spatial mask, only after deciding the color representation and the validity of the inputs. The arithmetic works for negative values as well; actual display values may require further handling.
A finite analysis and synthesis pair
Definition 1.25.1
(Reduction and expansion operators)
Let and be two aligned four-pixel signals. Define reduction and expansion by
The reduction averages before taking one value per pair; expansion repeats each pair value. These operators are explicitly defined on four pixels, so there is no hidden boundary rule. This crude box filter is chosen for hand calculation, not as a recommendation for high-quality photography. Its frequency response does not remove every frequency that can alias on downsampling.
Definition 1.25.2
(Coarse image and residual band)
Set and , with analogous definitions for . Call the coarse image and the residual or fine band. The residual is signed: it is not itself a displayable photograph. Reconstruction is exactly. This identity holds by definition even if reduction discards information: the discarded information is retained in .
Definition 1.25.3
(Masked blend of two bands)
For a fine mask with , put . At each level the mask weights image , and its complement weights :
Here means coordinatewise multiplication, and each denotes the all-ones vector of the relevant size. Since pair averaging preserves constants, . A mask of one selects at that location and zero selects at the corresponding level. Intermediate values cross-fade the two bands.
Theorem 1.25.4
(Identity and constant-mask properties)
For the operators above, if , the result is exactly for every mask. If every mask entry equals a constant , the result is exactly . All-zero and all-one masks are special cases.
Proof
(Identical images and constant masks)
If , then and . At each coordinate the weights and sum to one, and so do the coarse weights and . Hence by the definition of the residual.
If , then , because averaging a pair of equal entries returns that entry. The formula becomes , and linearity of gives
The cases and are the all-zero and all-one masks.
Proof X-ray
Where the constant-preserving reduction is needed
Once the mask is fixed, the blend is linear in the pair of images, so both statements compare weights band by band. The first statement uses only that equal inputs give equal bands and that complementary weights sum to one at each level. The second needs one more fact, . Suppose the coarse mask were chosen independently of the fine mask, say but . Then . For and this equals , which is neither nor , although the fine mask selects everywhere. The argument uses neither statistical independence nor a claim that the bands are perfectly separated in frequency.
Work through a seam
Worked example 1.25.5
(A broad base with a narrow detail switch)
Take , and . Both images are constant on each pair, so , , and . The coarse mask is and . Here the selected halves meet without a jump, but this is a property of these numbers, not a universal guarantee.
Worked example 1.25.6
(A single bright pixel crosses the mask boundary)
Use , and . Then , and . The fine contribution is and the expanded coarse contribution is , giving . Ordinary pixelwise blending would instead give . The coarse contribution crosses the fine-mask boundary: that is the intended wider transition at this level.
Worked example 1.25.7
(A seam inside a pair of pixels)
Let and , so and . For the pair-aligned mask we get and . This edge case keeps a hard step: the seam falls between two pairs, so the coarse level adds no transition. Move the seam to . Then and , whereas pixelwise blending gives . The pair containing the seam receives the intermediate value at both pixels instead of the values and .
Counterexample mode
A pyramid blend need not preserve the input range
Use , and . Then and , while . The fine term is ; the expanded coarse term is . Thus , despite both inputs and the mask lying between zero and one. Each band is mixed with convex weights, but the sum of signed bands is not necessarily a convex combination of the original pixels. Clipping would remove the negative value but changes the result and may alter local contrast.
From the example to an image pyramid
For a two-dimensional image, specify a blur, boundary extension, decimation phase and expansion rule at each level. Define , and for levels . Reconstruct from using . Backward substitution proves at every level, provided the same operators and array sizes are used in both directions. It does not require to be the identity.
Common mistake
(Expecting exact reconstruction to need a lossless reduction)
It is tempting to think that the pyramid reconstructs the image only if reduction followed by expansion loses nothing. It does not. Each residual is defined as , so whatever the reduction discards is stored in . For the reduction cannot be inverted, yet holds exactly. What a lossy reduction costs is the interpretation of the bands, not the identity.
Build a mask pyramid with the same stated reduction, blend the two residuals at each level with that level's mask and its complement, blend the final coarse images, then reconstruct. A normalized nonnegative blur makes reduced masks remain in ; this depends on the stated boundary treatment as well. Typical smooth filters and overlapping support avoid the blocky pair boundaries of our toy operators. Filter width, number of levels, alignment and mask shape remain design choices; two implementations with different choices need not give the same pixels.
A full image pair often has nonoverlapping regions.
Common mistake
(Treating an unobserved pixel as a black measurement)
Do not treat an unobserved pixel as an ordinary black measurement: filtering it into valid regions creates dark fringes. A measured zero is data, while an unobserved pixel is unknown.
Use valid samples with a stated normalization, or extend each image by a stated method before building the pyramid, and restrict the result to the union of valid support. A normalized masked blur divides weighted image sums by the corresponding positive sum of validity weights; if that denominator is zero, the value is unsupported. Its nonlinear, spatially varying operation requires separate analysis and does not automatically inherit every identity above.
Concept lensStructural
What blending can and cannot decide
A mask encodes which aligned source to favor. If two photographs show a moving person in different positions, blending does not determine the true position: a broad fine-scale transition can produce two translucent people. Registration, motion-aware source selection and seam placement should be decided using the actual data; multiscale reconstruction addresses a different problem.
Practice with solutions
Quick check
(For , calculate and , and reconstruct .)
Solution · Answer
and , so . Adding the expansion back yields exactly.
Quick check
(For , calculate both and .)
Solution · Answer
Both are . More generally , by linearity and constant preservation. This equality need not hold for a reduction that does not preserve constants.
Quick check
(Suppose and the mask is . Find the output without evaluating every band.)
Solution · Answer
The identity theorem gives . The masks select between identical residuals and identical coarse images; their complementary weights add to one.
Quick check
(Does the range counterexample imply that every output built from a negative residual is negative? Does treating an unobserved input as zero preserve the same meaning as a measured zero?)
Solution · Answer
No to both. A negative residual can be canceled by its positive coarse reconstruction, as in the reconstruction check above. The counterexample only shows that range preservation is not guaranteed after unequal band masks. A measured zero is data, while an unobserved pixel is unknown; inserting zero silently changes filtered values near the missing region.