ROI vs binning vs subsampling: choosing faster acquisition

When image acquisition cannot keep up with production, sending fewer pixels is an obvious option. The difficult part is deciding which pixels you can afford to lose. Cropping, combining pixels and skipping samples can produce identical output dimensions while preserving very different inspection information.

Choose sensor ROI when you can discard part of the scene without losing inspection coverage. Consider binning when you need the original field but can accept coarser sampling. Treat subsampling cautiously when small defects or repeating textures matter. This comparison shows how to calculate the sampling margin, verify the camera implementation and measure the actual throughput improvement.

Choose by coverage and defect size first

A region of interest, or ROI, selects a smaller rectangular area. With unchanged optics and native pixel sampling, the retained area keeps its original object-space detail. However, the image no longer covers the entire original scene.

Binning combines multiple sensor samples into fewer output samples. It can retain approximately the original field of view, but reduces the number of independent spatial measurements across that field. Edge alignment and supported sensor modes may trim the boundaries.

Subsampling, sometimes called decimation, selects samples while omitting others. It can also retain approximately the original field, but fine detail may disappear or produce misleading patterns.

The practical decision is not which setting produces the smallest file. It is which setting preserves the required detection performance while removing the actual bottleneck.

For broader exposure and tonal adjustments, use the industrial camera image quality settings guide. Those controls do not replace a sampling budget.

Calculate the pixels required for the smallest defect

Start with the smallest relevant defect dimension, not the camera's megapixel label. For a fixed field of view, horizontal sampling follows this relationship:

Use consistent length units throughout. Repeat the calculation vertically when defects can have different orientations or dimensions.

Suppose the required field is 160 mm wide and the smallest defect is 0.20 mm wide. Assume your inspection requires at least 4 samples across that defect. This is an application requirement to validate, not a universal detection threshold.

A hypothetical image with 4000 horizontal pixels provides:

A factor-of-two reduction horizontally leaves 2000 samples across the same field. Sampling becomes 0.080 mm/pixel, so the defect spans only 2.5 samples. That fails the assumed requirement even though the defect might remain visible in a favorable preview.

The camera resolution calculator helps establish the initial pixel requirement. Its calculation is a starting point; lens contrast, defocus, noise and defect orientation still affect detection.

Work through equal output sizes with different information

Continue with a hypothetical 4000 × 3000 image covering 160 × 120 mm. Compare three ways to produce a 2000 × 1500 output.

A native-sampling ROI retains half the width and half the height. With unchanged optics, it covers 80 × 60 mm. Sampling remains 0.040 mm/pixel, and the 0.20 mm defect still spans 5 samples inside the retained area.

Binning by two along each axis can retain approximately 160 × 120 mm while producing the same output dimensions. Sampling becomes 0.080 mm/pixel, leaving only 2.5 samples across the defect.

Subsampling by two along each axis has the same nominal output spacing as that binning example. However, omitted samples make its response to narrow defects more sensitive to feature position and texture.

For this example, ROI is the preferred option only if the smaller field contains every required inspection location, including placement variation. Neither reduced-sampling option meets the assumed defect requirement across the original field.

Do not compensate by changing the lens without recalculating coverage. Magnifying the part to recover defect sampling also sacrifices field of view.

Distinguish sensor readout from output processing

A camera can expose a feature named ROI without shortening sensor readout. The crop might occur after the sensor has delivered its full image. In that case, transmission and host processing can improve while the sensor-limited acquisition rate stays unchanged.

Read the manufacturer's feature documentation for these distinctions:

• Sensor windowing: whether the sensor reads fewer rows, columns or both.
• Camera-side cropping: whether pixels are removed after sensor readout but before transmission.
• Binning location: whether combination happens in the sensor, camera electronics or host software.
• Decimation behavior: which samples are retained and how offsets affect their positions.
• Mode dependencies: whether pixel format, shutter mode or another setting restricts the feature.

Reducing image height often affects readout differently from reducing width because many sensors operate through row-based timing. That is a reason to test both dimensions separately, not a promise of a particular speed increase.

On smart cameras, an inspection-tool ROI may limit algorithm processing without changing image acquisition. Ask whether the setting changes sensor readout, transmitted output, inspection workload or only the displayed region.

When a feature is absent or unavailable, follow the GenICam feature access troubleshooting guide. A similarly named control on another camera does not establish support on yours.

Understand binning noise and subsampling aliasing

Binning is not restricted to one sensor technology. Its behavior depends on the implementation, including whether values are combined before or after individual samples are read.

Combining samples can improve the signal-to-noise ratio of a larger effective sampling area. For 4 independent samples with equal signal and uncorrelated noise, summing produces 4 times the signal and 2 times the noise. The signal-to-noise ratio therefore improves by a factor of 2 under those assumptions.

That calculation does not guarantee better defect detection. A narrow dark defect combined with several bright background samples can lose contrast. Correlated noise, clipping, scaling and the readout architecture also change the outcome.

Subsampling does not combine the omitted samples into the output. Fine repeating structures can therefore appear as false coarse patterns. A narrow defect may also fall between retained positions and become weak or absent.

Binning provides some spatial averaging, but a simple averaging operation is not an ideal antialiasing filter. Test both methods against fine textures and edges near the inspection limit. The antialiasing guide for downsampling and warping explains why reducing image dimensions requires attention to spatial frequencies.

Color cameras need additional checks. A color-filter mosaic cannot always be combined or skipped like a monochrome grid. Verify supported color modes, output pixel format, color reconstruction and any restrictions on horizontal or vertical factors.

Estimate bandwidth, then measure the real rate

For an uncompressed image with a known transmitted storage format, payload demand can be estimated as follows:

Packed formats, padding and transport overhead require adjustments. Use the actual transmitted format rather than assuming sensor bit depth equals storage bits per pixel.

For the hypothetical 4000 × 3000 image transmitted at 8 bits per pixel:

A 2000 × 1500 output contains 3 MB per frame under the same assumptions. At 20 frames/s, its payload is 60 MB/s.

If a hypothetical system offers 120 MB/s of usable payload capacity, the bandwidth-only ceilings are 10 frames/s and 40 frames/s respectively. These are estimates, not camera performance claims. Exposure timing, sensor readout, trigger spacing and processing may impose lower limits.

The pixel format and exposure frame-rate guide covers other constraints that can prevent smaller images from increasing acquisition speed.

Measure distinct quantities instead of reporting one ambiguous frame-rate number:

1. Camera acquisition rate, using documented camera timing or counters where available.
2. Complete frames received by the acquisition application.
3. Frames processed into inspection decisions.
4. Trigger-to-decision latency and dropped or incomplete frames.

A faster display preview proves none of these by itself.

Verify each mode with a controlled test

Begin with a full-resolution baseline using fixed exposure, gain, lighting and pixel format. Keep the same part motion and inspection algorithm throughout the comparison.

1. Record geometry. Save width, height, offsets, sampling factors and any coordinate scaling used by inspection tools.
2. Change one dimension first. Test height-only ROI and width-only ROI separately before testing their combination.
3. Test binning independently. Restore the baseline before applying each supported factor and mode.
4. Test subsampling independently. Include translated fine features so retained-sample alignment changes between captures.
5. Measure sustained operation. Run each configuration for 60 seconds, repeat 3 times and record completed frames, timing variation and losses.
6. Repeat production decisions. Evaluate known acceptable parts and representative defects at the center, edges and expected placement extremes.
7. Restart and read back. Confirm that the intended acquisition configuration survives the actual startup sequence.

If changes require stopping acquisition, use the documented sequence rather than repeatedly writing unavailable controls. The guide to applying camera settings while streaming explains safe configuration boundaries.

Use the startup settings and readback checklist when a mode appears correct during setup but changes after reconnecting.

Update calibration and inspection coordinates whenever the image origin or scale changes. An ROI shifts the origin; binning and subsampling change the output sampling scale. Software that assumes the original coordinates can inspect the wrong location despite receiving valid images.

Select hardware without assuming feature support

Use the area scan cameras category and camera finder to narrow interface, shutter and resolution requirements. Then request the feature documentation for the exact candidate and operating mode.

The LS-CMR-0053 is a GigE rolling-shutter candidate. Its catalog information does not establish ROI, binning or subsampling behavior.

The LS-CMR-0166 is a USB3 rolling-shutter candidate. Confirm its supported acquisition modes independently rather than inferring them from its interface.

Before selecting either candidate, establish the supported sampling factors, ROI increments, offset restrictions and frame-rate behavior. Request an example configuration or documented timing information for the intended output dimensions.

For on-device inspection, also confirm whether tools operate on the original sensor image or a reduced image. Output dimensions alone cannot answer that question.

Final checks and when reduction is the wrong answer

Use this checklist before committing to a reduced-data acquisition mode:

• Coverage: the selected field includes placement variation and every inspection region.
• Sampling: the smallest defect retains the validated number of samples in both axes.
• Contrast: small defects remain distinguishable after combination or sample omission.
• Aliasing: fine textures do not create false features or hide real ones.
• Throughput: sustained completed-frame and decision rates meet production requirements.
• Coordinates: calibration, overlays and inspection regions match the new geometry.
• Startup: readback confirms the intended mode after reconnecting and restarting.

Keep native acquisition when the full field and smallest-defect sampling are both necessary. If readout is already faster than exposure permits, reducing pixels may not raise acquisition rate. If processing is the bottleneck, optimizing the algorithm or its processing ROI may preserve more inspection information.

For the worked example, native-sampling ROI is appropriate only when reduced coverage is acceptable. If full coverage is mandatory, retain sufficient native pixels and address transmission or processing capacity instead.

Start with the camera configurator, or send your field size, smallest defect, required decision rate and proposed sampling mode through applications support. The next decision should be based on a documented mode and a measured inspection margin, not output dimensions alone.