How to measure barrel distortion correction accuracy

A corrected image can look convincing while still giving the wrong dimensions near its edges. This matters when you are comparing correction settings, commissioning an inspection station, or deciding whether a calibration is ready for measurement. The useful comparison is against known geometry, not against the appearance of the distorted input. Follow the workflow below to separate straightness, camera-model accuracy, and actual measurement performance, then select parameters using evidence that was not used to fit them.

Choose geometric tests, not image correlation

Barrel distortion correction accuracy has several meanings. A correction can straighten a line without recovering the correct scale, and a low calibration fitting error can hide poor performance on new images.

Use these complementary tests:

Image correlation measures similarity of intensity patterns. Correcting distortion deliberately moves those patterns, while interpolation and cropping change them further. A nearly unchanged image can therefore correlate strongly with the input while retaining substantial geometric error.

Keep correlation for tasks where intensity similarity is relevant. Do not maximize it to choose distortion parameters.

If you first need to identify the model and coefficient convention, read interpreting distortion coefficients in an unrectified image. Coefficients from different conventions are not interchangeable.

1. Freeze the pipeline and separate your datasets

Before comparing correction settings, make the experiment repeatable:

1. Fix the camera position, lens focus, aperture, working distance, and target support.
2. Record image dimensions, sensor region, binning, exposure, and any software resizing.
3. Define the corrected image dimensions, output camera matrix, crop, and interpolation method.
4. Divide target views into fitting, validation, and final-test sets before tuning parameters.
5. Save the original images, detected reference points, parameter files, and software configuration.

The fitting set estimates the model. The validation set helps choose its settings. The final-test set supplies the acceptance result and must remain unused during tuning.

For example, an initial experiment might allocate 18 distinct views as 12 fitting views, 3 validation views, and 3 final-test views. This is an allocation example, not a sufficient-view guarantee. Coverage, target tilt, focus quality, and pose diversity matter more than a fixed count.

Include target observations near the image boundaries and corners. Several nearly identical central views provide little evidence about correction at the edges.

Use calibration target coverage margins to plan placement. Check target flatness requirements before treating a printed target as a geometric reference.

2. Measure straight-line residuals across the field

Place a verified straight edge or suitable target line at several positions and orientations. Include central, intermediate, and peripheral locations rather than relying on one horizontal edge.

A physical straight line projects to a straight image line under an ideal pinhole model, even when perspective changes its angle. However, a bent target edge is not a valid reference.

For each candidate correction, transform the same detected source points into corrected coordinates. Fit a line using orthogonal distances, rather than ordinary vertical residuals, so steep and shallow lines receive comparable treatment.

Here, u and v are corrected pixel coordinates, and r_i is the perpendicular point-to-line distance in pixels.

Report RMS, a high percentile such as the 95th percentile, and the maximum residual. Keep results grouped by line and image region. A single pooled average can conceal one badly corrected corner.

Use the same point-selection and outlier policy for every candidate. Repeatedly detecting edges after resampling can mix geometric correction quality with interpolation-dependent localization changes. A second test on the final resampled image is useful, but label it separately.

Straightness is a diagnostic, not a complete calibration. An image can contain straight lines while having the wrong scale, aspect ratio, or perspective transformation for your measurement task.

3. Test reprojection on views excluded from fitting

A reprojection test compares observed target points with locations predicted by the camera model. Low error on fitting images alone does not demonstrate that the model generalizes.

For each held-out view, keep the fitted intrinsic and distortion parameters fixed. Estimate only the target pose needed for that view. Where practical, use a spatially distributed subset of points for pose estimation and different points for scoring. Points used to estimate pose are not fully independent test observations.

For the primary camera-model score, compare observed points in the original image with forward projections into that same distorted image coordinate system:

This native-coordinate score avoids making a candidate look better merely by shrinking its corrected output. If you also score corrected coordinates, use the output camera matrix associated with that correction map and state the output scale.

Report results per view and per image region. An edge-heavy view with poor results should not disappear inside a large collection of accurate central points.

Look at residual direction as well as magnitude. Repeated spatial patterns suggest systematic model or setup errors; scattered residuals can also reflect target detection noise. Neither pattern identifies a cause by itself.

A planar transformation fitted to each corrected target image is not a replacement for this test. It can absorb scale, orientation, and other differences without proving that the camera model is correct.

If calibration becomes unstable when you add parameters, review diagnosing singular calibration estimates before increasing model complexity further.

4. Measure independent distances at the inspection plane

The most useful acceptance test usually measures what the inspection actually needs. Use a reference with independently known dimensions and uncertainty appropriate to your tolerance.

Measure distances near the center, edges, and corners, with horizontal, vertical, and diagonal orientations where relevant. Keep the reference at the intended measurement plane. Objects at different depths generally require additional geometry; distortion correction alone does not remove perspective-related scale changes.

Suppose an independently established distance is 50.00 mm, and your complete corrected-image measurement returns 49.94 mm:

Repeat the measurement after repositioning the reference, not just by processing the same image repeatedly. Record both systematic bias and variation between acquisitions.

Pixel residuals can help interpret the result, but their physical meaning depends on local scale. For an approximately uniform planar mapping spanning 160 mm across 2000 pixels:

That displacement is not automatically a length-error bound. Length measurements involve multiple localized features, and perspective can make scale vary across the image.

If the sampling requirement is 0.05 mm per pixel across 160 mm, the minimum image width from sampling alone is 160 / 0.05 = 3200 pixels. Use the calculator to check this separate acquisition requirement.

For the conversion from image coordinates to physical dimensions, use length measurement from image markers. Keep its metric-calibration checks separate from the distortion-model checks above.

5. Choose correction parameters using validation results

Set acceptance limits before inspecting the final-test results. Limits should follow the inspection requirement, reference uncertainty, and observed feature-localization repeatability—not a universal pixel threshold.

The following values illustrate how competing candidates can behave. They are not product specifications or recommended acceptance limits.

Candidate C gives the straightest lines, but candidate B performs better on the other geometric checks. If the application requires maximum absolute length error below 0.10 mm, only B passes that particular validation requirement.

This does not establish final acceptance. Freeze the selected candidate and run the untouched final-test set. If you change parameters after seeing that result, the set has become validation data; collect new final-test evidence.

Prefer the simplest model that meets the requirement across the usable field. Extra coefficients can reduce fitting error while making predictions unstable outside well-observed regions.

A single correction-strength control may be inadequate when the assumed distortion center is wrong or the model cannot represent the lens. For very wide-angle imaging, check whether the chosen projection model is appropriate instead of forcing increasingly extreme coefficients into an unsuitable one.

6. Verify the delivered image and its limits

A good fitted model can still produce a bad delivered image if map generation or resampling is wrong. Verify the exact processing path used by the inspection application.

Check the following separately:

• Map direction: confirm whether the remapping function expects output-to-input sampling coordinates or forward-transformed source coordinates.
• Coordinate conventions: check pixel-center definitions, axis order, units, and coefficient ordering.
• Output geometry: use the output camera matrix and crop offset when interpreting corrected coordinates.
• Valid image area: exclude blank borders and regions that sample outside the source image.
• Interpolation: inspect how resampling affects edge localization, especially near the periphery.
• Repeated acquisition: capture new images after a restart and confirm the saved configuration is applied.

A useful software check is to transform test points forward and backward through the mathematical mapping. Small round-trip error verifies numerical consistency, not real-world accuracy. A consistently wrong model can still pass that check.

Also test motion if the station measures moving parts. Rolling-shutter deformation, motion blur, target bending, and changing depth are not ordinary static barrel distortion. A correction fitted to a stationary target does not establish accuracy under those conditions.

Record any excluded border region explicitly. Cropping away a poor corner can be a legitimate design decision, but the resulting smaller usable field must still cover the inspection task.

Check hardware and save an acceptance record

Do not replace hardware solely because two corrected images look similar. First identify whether the limitation is reference quality, model choice, sampling, focus, motion, or the usable field.

If optics need changing, compare coverage, mount, working distance, and required field using machine vision lenses and the lens finder. The focal-length selection guide helps when the field or working distance must change.

The LS-FFL-0602 is a C-mount machine vision lens to evaluate against those requirements. Its catalog listing does not establish a correction-accuracy guarantee.

For acquisition changes, compare area scan cameras through the camera finder. Select acquisition settings and shutter behavior for the actual inspection conditions.

The LS-CMR-0053 is a GigE area scan camera with a rolling shutter. Evaluate motion separately before treating a static calibration result as a production result.

Before accepting the system, save this short checklist with its results:

• Independent fitting, validation, and final-test image lists.
• Camera settings, lens settings, model convention, and correction-map configuration.
• Straight-line and reprojection residuals, separated by image region.
• Known-distance errors, repeatability, and reference uncertainty.
• Usable output boundaries and excluded pixels.
• Acceptance limits and the final pass or fail decision.

Start a revised optical setup with the lens configurator, or bring the acceptance record and representative images to application support for help identifying the limiting factor.