Matlab Process Images

Preprocessing

Diagnostic decision guide for preparing images before processing: data type interpretation, conversion, normalization, contrast enhancement, illumination correction, and arithmetic. Start from the problem (dark image, uneven lighting, wrong range), not the function name.

Often loaded with: filtering.md (smoothing/denoising), image-io.md (loading data), deep-learning-imaging.md (per-channel normalization for networks).

Routes to other references: color space conversion → color-processing.md; morphological illumination correction (top-hat, bottom-hat) → morphological-ops.md; indexed image semantics → image-io.md; volumetric preprocessing → 3d-volume-processing.md; display viewers → skill matlab-display-image.

Characterize Before Transforming

Before any preprocessing step, establish these properties (for full characterization workflows, see image-understanding.md):

Key insight: conversion is not normalization. im2single does type-aware scaling (uint8 255 → 1.0); rescale remaps actual min/max. double(img) does neither — it just changes the type without scaling, producing values like 128.0 that IPT functions interpret as out-of-range.

Problem-Oriented Preprocessing Workflow

1. Diagnose — What is wrong? (dark, low contrast, uneven illumination, wrong type for downstream task)

2. Choose one transformation — Pick the minimal operation that addresses the problem

3. Inspect the result — Verify the fix didn't introduce new issues

4. Evaluate — Does it improve the downstream task? (see quality-assessment.md)

5. Only then add another stage — stacked operations obscure which step caused a failure

Typical Preprocessing Order

When multiple preprocessing steps are needed, this order avoids common pitfalls:

1. Illumination correction (imflatfield, imtophat) — before contrast enhancement, which would amplify the illumination gradient

2. Denoising (imgaussfilt, medfilt2, imnlmfilt) — before contrast enhancement, which would amplify noise

3. Contrast enhancement (imadjust, adapthisteq) — after illumination and noise are addressed

4. Type conversion (im2single, im2uint8) — at the boundary where downstream functions require a specific type

Illumination correction, contrast enhancement, and histogram equalization are covered in the sections that follow this order. Denoising is covered in filtering.md. The remaining sections — image arithmetic, linear combinations, and image blending — cover operations typically applied at or after the type-conversion boundary.

Display vs Data — Know the Difference

Adjusting the viewer's display range changes presentation only. Calling imadjust or rescale changes the actual pixel data. Before modifying data, ask: does the output need to be saved or only inspected? For scientific/medical data, prefer display-range adjustment over data modification when the goal is visualization.

% Display-only: adjust viewer window/level — data is untouched
imageshow(img, DisplayRange=[500 2000]);

% Data modification: pixels are permanently changed
adjusted = imadjust(img, [500/65535 2000/65535], []);

Illumination Correction

imflatfield

corrected = imflatfield(img, sigma);

• sigma should be larger than the largest foreground object (typical: 30–100)

• Supports a Mask argument to limit where the correction is applied — pixels where the mask is false retain their original values

• Processes RGB images through the HSV value channel automatically

• Simpler than morphological alternatives for smooth illumination gradients

Contrast Enhancement — Choosing a Method

Histogram Equalization

Color images: Never apply histeq or adapthisteq directly to RGB — per-channel equalization distorts color. Convert to L*a*b* or HSV and equalize only the luminance/value channel.

imadjust Details

adj = imadjust(gray);                           % Auto 1% saturation
adj = imadjust(rgb, stretchlim(rgb), []);        % RGB — required form
adj = imadjust(img, [0.2 0.8], [0 1]);           % Explicit input range
adj = imadjust(img, [], [], 0.5);                 % Gamma correction

• Single [low high] applies to all channels; a 2-by-3 matrix allows per-channel limits

CLAHE — adapthisteq

enhanced = adapthisteq(img, ClipLimit=0.02, NumTiles=[8 8]);

• ClipLimit (0–1): higher = more contrast; lower = prevents noise amplification. Start with 0.02

• NumTiles: more tiles = more local; can expose tile-boundary artifacts at high values

• NBins: trades speed for dynamic-range fidelity (default 256)

• Grayscale only — for color images, apply to luminance channel only (convert to L*a*b*, enhance L, convert back, clip out-of-gamut). See color-processing.md for the full pattern

Histogram Matching — imhistmatch / imhistmatchn

Transform an image so its histogram approximates a reference image's histogram. Use for batch normalization, cross-session consistency, or matching imaging conditions:

matched = imhistmatch(img, referenceImg);           % 2-D grayscale or RGB
matched = imhistmatchn(vol, referenceVol);           % N-D grayscale

• imhistmatch accepts grayscale or RGB (per-channel matching for RGB)

• imhistmatchn is grayscale only, works on N-D data (volumes)

• Both accept uint8, uint16, int16, single, double

Conversion and Normalization Decision Tree

Do NOT use double(img) for image conversion — it changes the type without scaling, leaving values like 128.0 that IPT treats as out-of-range.

Quantization warning: converting from higher to lower bit depth (e.g., uint16 → uint8) permanently discards precision. Do this only at the end of a pipeline.

Batch warning: rescale and imadjust with default arguments normalize each image to its own min/max. In a dataset where intensity comparisons across images matter (time series, dose response, multi-well plates), compute normalization limits once from the entire dataset and apply them uniformly — otherwise inter-image relationships are destroyed.

Prefer im2single over im2double:

• Halves memory (4 bytes/pixel vs 8)

• Many IPT functions have optimized code paths for single

• 24-bit mantissa covers uint16 range exactly

Most IPT functions accept integer images directly — do not convert to floating-point as a default habit. If an IPT function rejects your data type, check its documentation with matlab-read-documentation — the issue may be a specific type restriction (e.g., int16 not supported), not a blanket float requirement.

When float IS needed: custom arithmetic operations (matrix math, weighted sums — integer clips silently), deep learning network input (see deep-learning-imaging.md for per-channel normalization with ImageNet statistics and dataset mean/std).

Image Arithmetic and Precision

Integer arithmetic clips silently. Nested operations compound rounding error:

Precision rule: If a computation involves subtraction, weighting, or multiple steps, convert to single once at the start, compute, then cast to the output type once at the end:

% WRONG — integer clips at each step
result = imadd(imsubtract(uint8(A), uint8(B)), uint8(C));  % Lost negatives, compounded rounding

% RIGHT — compute in float, cast once
A = im2single(A); B = im2single(B); C = im2single(C);
result = im2uint8(A - B + C);

imlincomb — Linear Combinations Without Intermediate Clipping

For weighted sums, imlincomb computes the entire expression in double precision and rounds only at the end. Use imlincomb when you need precise weighted arithmetic on integer images and the result should remain integer. For visual compositing and blending effects, use imblend instead.

% Weighted blend — no intermediate clipping
result = imlincomb(0.7, img1, 0.3, img2);           % Returns same class as img1
result = imlincomb(0.5, A, 0.5, B, "single");       % Force output class
result = imlincomb(1, img1, -1, img2, 128);          % Difference with offset

Prefer imlincomb over manual a*im2single(A) + b*im2single(B) when the output should remain integer.

Image Blending — imblend

Blend two images using various modes (R2024b). Do NOT manually implement alpha blending:

blended = imblend(foreground, background);                        % Alpha blend (default)
blended = imblend(fg, bg, mask, Mode="Poisson");                  % Seamless Poisson blend
blended = imblend(fg, bg, mask, Mode="PoissonMixGradients");      % Mix gradients at boundary
blended = imblend(fg, bg, mask, Location=[100 50]);               % Place fg at offset in bg

Common Mistakes

Copyright 2026 The MathWorks, Inc.