focus_stack.analysisRun.computeFocusStack

Description

Builds a Laplacian pyramid for each input image, estimates local detail energy at every pyramid level, and blends pixels using the resulting focus weights. Low-confidence regions use normalized soft weights instead of a hard source selection. The first image defines the output height and width; other images are resized to that geometry. Grayscale and RGB inputs may be mixed, in which case the working stack and output are RGB.

This function fuses focus only. It does not align images, compensate for parallax, or correct moving subjects. Use the app's registration step before calling it when the camera or specimen moved between frames.

Inputs

images

At least two numeric images supplied as a cell array or numeric stack. A four-dimensional array uses its fourth dimension as the image index. A three-dimensional numeric array is interpreted as a stack of two-dimensional grayscale images, not as one RGB image; use a cell array for multiple RGB images. Values are converted with labkit.image.im2double.

opts

Optional scalar structure containing the fusion options below.

Options

focusWindow

Local detail-energy window in pixels. The value must be a positive finite scalar; it is rounded up to an odd integer of at least three. Default: 31.

smoothRadius

Radius in pixels used to smooth source weights at the finest pyramid level. It is rounded to a nonnegative integer and reduced at coarser levels. Default: 4.

minConfidence

Threshold in [0,1] below which fusion uses relative detail scores rather than only the best source image. Default: 0.05.

pyramidLevels

Requested number of Laplacian levels. It is rounded and limited to the range supported by the image geometry, from one through five. Default: 4.

Outputs

result

Scalar structure containing the fused image and diagnostics.

Result Fields

ok, message - true and empty text after successful fusion.

fused

Grayscale or RGB double image in [0,1].

focusIndex

M-by-N uint16 map naming the source with the largest finest- level detail score at each pixel. Values run from 1 to inputCount.

confidence

M-by-N double map in [0,1] measuring separation between the best and second-best finest-level detail scores.

focusCoverage

One fraction per source image, calculated from focusIndex; the fractions sum to one. inputCount, imageHeight, imageWidth, channelCount - Working stack geometry. focusWindow, smoothRadius, minConfidence, pyramidLevels - Effective options.

meanConfidence

Mean of the confidence map.

method

"Laplacian pyramid focus fusion".

resizedCount

Number of input images resized to the first image's geometry.

Errors

labkit_FocusStack_app:NotEnoughImages

Fewer than two images are supplied.

labkit_FocusStack_app:InvalidImages

The input container or one of its elements is not a supported numeric image. MATLAB validation errors - focusWindow or minConfidence is outside its documented numeric domain.

Example

[x, y] = meshgrid(1:24, 1:20);
first = double(mod(x, 4) < 2);
second = double(mod(y, 4) < 2);
result = focus_stack.analysisRun.computeFocusStack( ...
{first, second}, struct("focusWindow", 5, "smoothRadius", 1));
assert(result.ok && isequal(size(result.fused), [20 24]))
assert(abs(sum(result.focusCoverage) - 1) < 1e-12)