Machine vision terminology: from capture to measurement

A machine vision inspection can fail because a camera captured the wrong moment, an image-processing tool selected the wrong feature, or a measurement used the wrong scale. Those are different problems, even when they appear in the same software window. This vocabulary guide follows an inspection from capture to dimensional measurement, explaining what each operation changes, what it leaves unchanged, and which questions to ask before selecting hardware.

Follow the inspection pipeline, not the menu order

Machine vision turns images into measurements or decisions that another system can use. A typical inspection captures an image, isolates relevant information, extracts features, and compares results against acceptance limits. The final output might be a diameter, a defect location, or a pass/fail signal.

The stages are related, but they are not interchangeable:

These stages need not occupy separate software tools. One measurement tool might perform filtering, edge detection and fitting internally. Ask which operations it applies before adding duplicate processing upstream.

A useful diagnostic rule is to inspect the output of each stage. If the original image lacks the feature, changing the final acceptance tolerance cannot recover it. Conversely, a sharp image does not prove that the measurement is calibrated correctly.

Acquisition terminology describes how images arrive

Acquisition means obtaining image data from a camera or another supported source. It does not necessarily mean continuous capture, and it does not guarantee that processing has started.

Free-running acquisition means the camera generates frames without an individual external command for every exposure. Triggered acquisition means a configured event requests a frame or acquisition sequence. That event may be an electrical input or a software command.

Grab, snapshot and single acquisition are software-dependent labels. Some applications use “grab” for continuous capture; others use it for retrieving any frame. Check whether a command starts a new exposure, reads an existing buffer, or starts a continuous stream.

A receiver can run continuously while the camera waits for triggers. Therefore, “continuous acquisition” in the application and “triggered exposure” in the camera are not necessarily contradictory settings.

The software-triggered capture guide explains the distinction between a user-interface action and an actual acquisition request. For electrical triggering, the hardware-trigger frame-rate guide covers the relationship between incoming pulses and accepted frames.

Other terms belong to the communication layer:

• Driver: software that allows the operating system or application to communicate with supported hardware.
• SDK: a software development kit containing interfaces, libraries and documentation for application development.
• Buffer: memory holding image data before or during processing.
• Frame ID or timestamp: metadata that can help identify frame order or timing, when provided by the acquisition system.
• Simulated acquisition: processing saved images instead of receiving new camera frames.

Saved-image tests are valuable for repeatable algorithm comparisons. They cannot validate exposure timing, trigger acceptance, cable reliability or live buffer handling. Obtain acquisition software and documentation through the manufacturer or its authorized distribution channels.

Camera settings are not image-processing operations

Exposure time controls how long the sensor collects light. Gain amplifies the signal; it does not collect additional photons. Pixel format describes how image samples are represented, including their organization and bit depth.

Global shutter and rolling shutter describe exposure timing across the sensor. A global shutter exposes the image together, while a rolling shutter exposes different rows at different times. Moving subjects can therefore produce different geometric artifacts even when individual rows look sharp.

A brightness adjustment applied after capture cannot remove motion blur or reconstruct clipped highlights. Similarly, a sharpening filter cannot reliably restore detail that the optical system never resolved.

Consider an illustrative inspection with an object-space scale of 0.05 mm per pixel. The part moves at 500 mm/s, and exposure lasts 0.1 ms:

This estimates translational blur during exposure. It excludes rolling-shutter geometry, vibration, defocus and transport timing errors. Treat it as an acquisition requirement, not a promised camera result.

An attribute, feature or node is a readable or writable camera setting or item of information. Some features become unavailable while acquisition runs. After changing a setting, read it back where supported and check the resulting image.

Use the exposure and gain units guide when software displays raw values rather than clearly labeled physical units. For repeatable startup behavior, follow the camera settings initialization guide.

Filtering and thresholding answer different questions

A filter transforms pixel values, usually using neighboring samples. Filtering may suppress noise, emphasize transitions or remove unwanted spatial detail. It does not automatically distinguish defects from acceptable surface texture.

Common examples include:

• Mean filter: averages neighboring values, reducing fluctuations while softening edges.
• Gaussian filter: applies a weighted average, with nearby samples generally contributing more strongly.
• Median filter: replaces a sample with the neighborhood median, often helping with isolated bright or dark specks.

Filter size is an inspection decision. A neighborhood spanning 5 pixels can alter a feature only 2 pixels wide. Compare filtered and original images at the smallest defect or edge you must preserve.

Thresholding classifies samples according to an intensity rule. For a grayscale image, a simple global threshold uses the same limit throughout the selected area:

For example, with an illustrative 8-bit image and a threshold of 120, an intensity of 119 enters one class and 120 enters the other. Display software may draw those classes as black and white, but the underlying logical values need not be stored as display intensities.

Adaptive thresholding varies the threshold according to local image information. It can help with uneven illumination, but its neighborhood size can also change which structures survive.

Color images require an explicit classification rule, such as thresholding a selected channel or a derived intensity image. Converting color to grayscale can remove distinctions between differently colored materials with similar brightness.

The binary preview guide is the closest practical follow-up for checking segmentation while images arrive. Compare acceptable and defective samples before deciding that one threshold is sufficient.

ROI, masks and edges define what gets measured

A region of interest, or ROI, limits an operation to a selected part of the image. A mask identifies included or excluded pixels, often allowing more complex shapes than a rectangular ROI.

Distinguish an acquisition ROI from a processing ROI. An acquisition ROI restricts the image region delivered by the camera, subject to device capabilities. A processing ROI limits analysis after the image arrives. Reducing a processing ROI may reduce computation without changing camera exposure, transmitted image dimensions or acquisition rate.

A third use is an automatic-control ROI, which selects the area used to calculate exposure or gain. Moving that region can change the image itself. The automatic-exposure ROI guide explains why linking control regions to unstable detections can create feedback problems.

An edge is a spatial change in image intensity. It may correspond to a physical boundary, but it may also come from a shadow, reflection, printed mark or surface texture.

Several terms describe what happens next:

• Edge polarity: the transition direction, such as dark-to-light along a defined search direction.
• Edge strength: a measure of intensity change, with the exact calculation depending on the tool.
• Contour: an ordered boundary representation extracted from image information.
• Line fit: a geometric line estimated from selected points.
• Circle fit: a circle estimated from selected points.
• Subpixel localization: an estimated position between pixel centers, based on a model of the sampled transition.

A circle fit can return a precise-looking diameter from the wrong contour. Restrict the search region, check polarity, inspect rejected points and overlay the fitted result. Subpixel output provides finer numerical resolution; it does not by itself establish dimensional accuracy.

Calibration connects image geometry to physical dimensions

Calibration establishes a relationship between image coordinates and a known reference. A simple scale factor may be adequate for a flat object, fixed working distance and sufficiently controlled geometry. Lens distortion or perspective can require a spatially varying mapping instead.

Calibration does not necessarily modify the displayed image. Software may retain the original pixels and transform only detected coordinates. Rectification or undistortion, by contrast, usually refers to a geometric image transformation.

Consider a hypothetical planar inspection. A known reference length of 50 mm spans 1,000 pixels:

Now suppose the inspection must cover a width of 100 mm and represent a 0.2 mm feature with 4 pixels:

Use the resolution calculator to explore this sampling relationship. Four pixels across a feature is an illustrative design target, not a universal detection or measurement guarantee.

Sampling, repeatability and accuracy are different quantities. If each detected edge has an assumed localization error bounded by 0.2 pixel, their separation could have a worst-case error of 0.4 pixel. At the example scale, that contribution is 0.4 × 0.05 mm = 0.02 mm. Reference uncertainty, distortion, part height and mounting changes remain additional contributors.

The calibrated length measurement guide explains how reference geometry supports dimensional results. Verify calibration using a separate known dimension, preferably at several image locations and representative part heights.

Match the processing location to the inspection

The same vocabulary applies across different hardware arrangements. The main purchasing distinction is where acquisition control, image processing and result communication occur.

Area scan cameras provide complete image frames for an external processing system. Use the camera finder to compare acquisition characteristics after defining image width, exposure requirements, shutter behavior and interface needs.

The LS-CMR-0053 and LS-CMR-0166 are catalog starting points for comparing external-processing camera arrangements. Their listings identify rolling shutters, which must be considered against the motion conditions discussed above.

Smart cameras perform inspection processing on the device. That architecture can reduce external integration, but the available tools, recipe handling and result interfaces still need verification.

Review the LS-SMC-0063 as a device-side inspection candidate. The catalog listing is not evidence that it implements every processing operation described here; confirm the required workflow before selection.

Vision controllers provide another location for processing and integration with cameras. Camera connections alone do not establish algorithm support, software licensing or electrical compatibility with machine controls.

The MV-VC2101-128G40-NN is a controller candidate whose listed basic configuration excludes I/O and light functions. Plan those interfaces separately rather than assuming the controller supplies them.

Before ordering, request confirmation of the supported acquisition software, required processing tools, license scope, result protocol and any external interfaces. A product category describes an architecture, not a complete inspection specification.

Check the vocabulary against a real inspection

Use this short checklist when reviewing a demonstration, quotation or software specification:

• Identify the source: distinguish live camera frames from saved-image simulation.
• Define the timing: state what requests exposure and how stale or missing frames are detected.
• Record acquisition settings: include exposure, gain, pixel format and shutter requirements.
• Name each ROI: distinguish camera readout, automatic-control and processing regions.
• Inspect intermediate outputs: compare the original image, filtered result, binary mask and detected geometry.
• Preserve coordinate meaning: account for cropping, resizing or rotation before converting positions to physical units.
• Validate the decision: test acceptable and defective samples, not only one favorable image.
• Check physical accuracy: compare measured dimensions against an independent reference under representative conditions.

The usual pipeline is not always the right one. A grayscale edge tool may measure a boundary without thresholding. A surface inspection may need texture analysis rather than a contour. Additional processing is useful only when it improves a defined inspection requirement.

Start a hardware shortlist with the camera configurator, or send representative images, timing requirements and measurement tolerances through application support. Define the required result first, then select the acquisition and processing tools that can demonstrate it.