You shrink an image to fit a processing budget, or you warp it to rectify a tilted camera, and the edges come back ragged: staircases along straight lines, moiré in fine textures, thin features flickering between frames. Changing the interpolation mode from bilinear to cubic does not fix it, and the reason is structural rather than a quality issue in any particular library. This article explains why point-sampled interpolation aliases, how much prefiltering a given scale factor needs, and how to build antialiased resize and warp steps into an inspection pipeline without giving up measurement accuracy.
Aliasing is missing prefiltering, not bad interpolation
A digital image can only represent spatial frequencies below half its sampling rate: the Nyquist limit. When you shrink an image by a factor s, the destination grid samples the scene s times more coarsely, so its Nyquist limit drops by the same factor. Any detail finer than the new limit - a 1-pixel line in a 4x reduction, weave in a plastic surface, a barcode module - now lies above the limit. A resampling filter that takes a single weighted sample per destination pixel folds that detail back into the band you can represent, as jagged edges and moiré instead of honest grey.
Interpolation names the weighting of that single sample. Bilinear weights the four nearest pixels, bicubic weights sixteen, Lanczos weights a wider window. All of them point-sample. A bicubic kernel has negative lobes that sharpen edges and add ringing, so it can look crisper and alias worse. The fix is not a better interpolation; it is to low-pass filter the image before sampling it on the coarser grid. That prefilter step is what antialiasing means in resampling.
The area-average rule and how much blur a scale factor needs
The ideal prefilter for a uniform reduction by s averages every source pixel that contributes to one destination pixel. That area-average kernel has a width of about s pixels, and a Gaussian of similar width is the practical stand-in because it is separable, smooth and has no negative lobes. A common working rule:
Downscale by 2, blur with sigma near 1 px; downscale by 4, sigma near 2 px; downscale by 10, sigma near 5 px. Then resize with bilinear or cubic interpolation - after a correct prefilter the interpolation choice matters little.
Worked example. You have a 12 MP image from a 4000 x 3000 sensor and your inspection algorithm runs comfortably on 1000 x 750. The factor is 4000 / 1000 = 4, so blur with a Gaussian of sigma 2 px (a kernel radius of 4-6 px covers it) and then resize to 1000 x 750. A 3 x 3 blur - roughly sigma 0.85 px - leaves every frequency between 0.85 and 2.0 cycles per destination pixel unfiltered, which is exactly the band where moiré lives. That is why a single 3 x 3 pass often satisfies the eye at 2x but fails at 4x and beyond.
Two refinements matter in practice. First, anisotropic scale factors need anisotropic blurs: a 4x reduction in x and 1x in y takes sigma 2 px horizontally and none vertically. Second, thin features are the casualty. A line 2 px wide in the source becomes half a pixel after a 4x reduction and the prefilter spreads it across neighbours; decide explicitly whether such features are inspection targets before choosing the reduction factor.
Warping compounds the problem
A perspective warp is a spatially varying scale. Where the homography compresses the image, many source pixels map to one destination pixel and the same aliasing mechanism applies; where it stretches, interpolation smooths anyway. A single global blur is wrong by construction: correct in the compressed region means over-blurred in the stretched one. Three workable approaches, in order of effort:
- Warp, then reduce. Rectify at a resolution high enough that the worst local compression is only 2x, then prefilter with sigma 1 px and resize to the working resolution. Simple, and adequate for most camera-mounting angles under roughly 20 degrees of tilt.
- Two-pass with intermediate pyramid. Halve repeatedly with area averaging (a Gaussian pyramid) until you are near the target size, then apply the warp at that intermediate resolution and finish with a small reduction. Each halving is a clean 2x antialiased step, and cumulative error stays small for factors up to 10x or more.
- Jacobian-adaptive filtering. Compute the local scale from the warp's Jacobian at each destination pixel and set the kernel width accordingly. This is what high-end resampling libraries do; it is worth it when you rectify strongly tilted planes, such as a camera looking along a conveyor at 45 degrees.
If your geometry work starts from a calibration, the same undistortion and rectification chain is covered in camera calibration for length measurement from image markers, and the residual geometry checks that catch a wrong homography before you blame the resampler are in camera extrinsic calibration accuracy beyond reprojection error.
What this means when choosing cameras and lenses
Antialiasing is mostly a software step, but the hardware decides how often you need it and how much detail it costs. Oversampling the feature - 5 to 10 pixels across the smallest defect you must detect - leaves room for a prefilter and a 2x to 4x software reduction with nothing lost. The area scan cameras in the shop range from under 3 MP to 29 MP and above, and the camera finder filters by resolution, shutter and interface; the camera configurator narrows the choice from the sensor up. If you are sizing the sensor to the part rather than to the algorithm, the relationship between resolution and the smallest measurable feature is worked through in choosing camera resolution for dimensional inspection, and the pixel-count arithmetic for a moving conveyor is in sorting conveyor camera selection: pixels, speed and blur.
The lens matters just as much, because a resampling prefilter cannot remove detail the optics never resolved cleanly. Match the machine vision lenses to the sensor format so the image circle covers the sensor, and check that the lens resolution rating meets or exceeds the sensor's pixel pitch demands. Sensor format mismatch is its own source of edge artefacts, explained plainly in lens sensor format: 1/1.8 inch, 2/3 inch and 1 inch lenses explained. For the focal-length side of the same decision, use the lens focal length calculator or the calculator below.
There is one case where the antialiasing happens before the light ever reaches the sensor: sensor binning. A 2 x 2 binned readout averages charge in the silicon, which is a genuine area average with no software step at all. It costs resolution, so it suits pipelines that need the noise and file-size benefits of a reduction but not an arbitrary warp. Binning factors are fixed integers, so it never replaces a rectification prefilter.
Verifying your resize and warp steps
A few minutes of testing separates a correct pipeline from one that will surprise you on real parts:
- Point the camera at a pattern of fine parallel lines or a resolution test chart and reduce by your working factor. Clean greyscale gradients with no beating mean the prefilter is wide enough.
- Warp a checkerboard and inspect edge profiles across a straight line: a monotonic ramp over 1-2 destination pixels is correct; overshoot, ringing or stair-steps mean the kernel is too narrow for the local scale.
- Run the pipeline on the worst part, not the nominal one. Fine surface texture on a matte plastic or woven part is where aliasing shows first.
- Confirm measurement repeatability after the change. A sub-pixel edge fit on a smooth ramp is typically more repeatable than the same fit on a 1-pixel transition.
If you build and debug pipelines offline before touching the line, the same images work well for testing resampling settings; the workflow is set out in how to practice machine vision with offline images.
When not to antialias
Prefiltering is not free, and there are cases where the standard advice is wrong:
- Barcode and datamatrix reading near the resolution limit. Modules 2 px wide in the source sit right at the new Nyquist limit after any reduction; a Gaussian prefilter can drop decode rate. Keep native resolution, or bin at acquisition instead.
- Thin-wire and edge-location gauging. If the measured feature is a 1-2 px line, widen the reduction factor or the sensor, not the blur.
- Display-only thumbnails. Even a mild 3 x 3 blur before resize is visually better than none at these scales, but area-average resampling alone is usually enough below 2x.
- Sub-pixel downscale factors near 1.0. Reducing by 1.1x needs a wide, weakly weighted kernel that behaves oddly; cropping to an integer factor or leaving the image alone is often the better call.
A short checklist before you commit a resampling step to production:
- [ ] Scale factor s computed for both axes; anisotropic if they differ
- [ ] Gaussian sigma ≈ s/2 per axis, or a genuine area-average resample
- [ ] Warp done before the final reduction, or with Jacobian-adaptive kernels for strong tilt
- [ ] Fine-pattern test image shows no moiré at the working scale
- [ ] Smallest inspection feature still spans at least 3 destination pixels
- [ ] Decode or measurement rate verified against the unfiltered baseline
Antialiasing a resize or warp is one Gaussian with a width tied to the scale factor, applied before the sampling grid coarsens. Get that width right and the interpolation mode stops mattering; get it wrong and no interpolation setting will save the edges. If you would rather sanity-check the whole chain - camera, lens, light, and the processing budget that dictates your reduction - the configurator walks the choices in order, and the team offers free expert support with every order. For more on the systems side, see the machine vision systems topic page.
Questions
Why do all interpolation modes give jagged edges when I shrink an image?
Interpolation modes describe how to weight neighbouring source pixels around a point sample. None of them average the full area of source pixels that collapse into one destination pixel, so detail finer than the new pixel grid folds into moiré and stair-stepped edges. Switching from bilinear to cubic changes sharpness slightly but cannot fix the missing prefilter.
How much blur should I apply before downsampling an image?
Blur sigma should be about half the downscale factor, because that is roughly the width of the area-average kernel for that factor. Downscaling by 4 needs a sigma near 2 px; downscaling by 2 needs sigma near 1 px. Use a slightly larger sigma if the source is noisy or contains fine periodic patterns, and a slightly smaller one if you must preserve thin lines.
Does antialiasing blur hurt measurement accuracy in a vision pipeline?
Sub-pixel measurement usually improves, because edges become smooth ramps that edge detectors and centroid methods fit reliably. What you lose is the ability to resolve features near the new pixel pitch, so choose the blur before choosing what to measure. If a feature is only 2-3 pixels wide after downsampling, keep more resolution instead of blurring harder.
Should I resize images before or after perspective warp?
Do the geometry first and build the blur into the resampling, or warp at an intermediate resolution. Blurring, then warping with a scale change, samples the blurred image unevenly across the frame: the effective blur differs between compressed and stretched regions. Adjust the prefilter by the local Jacobian scale if your library permits it, otherwise warp first and prefilter the second reduction separately.
Can a higher resolution camera avoid downsampling artifacts?
Often yes. Acquiring at full resolution from one of the area scan cameras and doing the reduction in software with a proper area-average resample gives you control over the filter. Sensor-side binning averages charge in the silicon and is cleaner still, though it is fixed at integer factors such as 2 x 2 and cannot match an arbitrary warp.