How to inspect internal features with machine vision

An internal defect is not automatically visible to a camera, even when the outside of the part produces a clear image. A useful inspection plan starts by separating accessible internal surfaces, features visible in transmission, and structures hidden behind opaque walls. Each requires a different approach.

For teams evaluating an inspection station, the first deliverable should be a defined acceptance test rather than a camera specification. The steps below help you describe the feature, establish optical access, estimate sampling requirements, and decide when another sensing method deserves evaluation.

1. Define the feature and the acceptance decision

Replace “inspect the inside” with a requirement that names a feature, its location, and the decision. Examples include detecting a burr inside a through-hole, measuring a visible pocket opening, locating a missing internal component, or finding a void beneath an opaque surface. These are different imaging problems.

Record the following before selecting hardware:

• Material: opacity, translucency, surface finish, and any variation between production batches.
• Geometry: overall dimensions, wall thickness, openings, recess depth, and permitted viewing directions.
• Defect: its physical form, minimum rejectable size, and expected location.
• Measurement: the dimension being evaluated, its datum, and the allowed tolerance.
• Production conditions: presentation repeatability, available inspection time, and whether the part can stop or rotate.
• Evidence: examples of acceptable parts, borderline parts, and confirmed defective parts.

A requirement such as “reject an obstructed passage” still needs clarification. Must the passage transmit light, pass a particular gauge, or maintain a minimum cross-sectional area? A bright opening can confirm an optical path without proving the entire passage meets its dimensional requirement.

For accessible small features, the small-part dimensional measurement guide explains the broader measurement setup. Keep defect detection and dimensional acceptance separate when they require different evidence.

2. Choose the sensing route from physical access

An ordinary camera records light reaching its sensor. It cannot reveal arbitrary structures behind opaque material. Establish how information about the feature will reach the imaging system before discussing resolution.

Treat this table as a screening tool, not a promise that every defect is detectable. For example, a thin discontinuity may be difficult to distinguish in a radiograph if its orientation produces little change along the imaging path.

Use qualified specialist providers when evaluating radiation-based inspection. Define the defect and acceptance criteria first; the specialist should establish the appropriate equipment, controls, and validation method.

3. Test optical access and lighting before algorithms

For an open cavity, draw the viewing rays from the feature to the lens. Check the entrance rim, nearby walls, and any tooling that could block those rays. A larger opening does not guarantee visibility of the bottom corners.

Move the sample through its expected orientation range while observing the raw image. If a defect disappears behind a wall, threshold adjustments cannot recover it. You need another view, different presentation, suitable probe optics, or another sensing route.

For an open passage, backlights can produce a strong silhouette. Use the lighting finder to shortlist geometries only after establishing the illuminated area and available mounting space. A diffuse source and a collimated source can produce different edge appearances, especially around thick parts and reflective walls.

The diffuse-versus-collimated backlighting comparison explains that choice in more detail. Neither arrangement makes a silhouette equivalent to a complete internal profile.

For visible cavity surfaces, reflections may be the dominant problem rather than insufficient brightness. The internal-surface glare guide describes why illumination direction matters when viewing reflective interiors.

4. Calculate sampling, then budget measurement error

Field of view must include the feature and its presentation variation. Increasing that field without increasing pixel count reduces the number of pixels representing each defect.

For a planar object viewed approximately square-on, nominal object-space sampling is:

These relationships describe sampling, not guaranteed accuracy or detection probability. Lens blur, focus, contrast, noise, and processing can all reduce usable information.

Consider a hypothetical inspection with a field width of 100 mm and a minimum visible defect width of 0.20 mm. Suppose the initial feasibility target is 5 pixels across that defect:

Five pixels is a starting assumption for this example, not a universal detection threshold. A low-contrast defect may require more sampling, better lighting, or both. Use the dimensional camera-selection guide to connect sampling requirements with practical camera choices.

Now consider a width measured between two edges. If each edge has a hypothetical localization uncertainty of ±0.5 pixel, their worst-case combined contribution is ±1 pixel. At the example sampling, that corresponds to ±0.040 mm.

If the allowed part tolerance is ±0.050 mm, this simple bound leaves only 0.010 mm before reaching the tolerance boundary. Calibration, presentation, focus, and other errors have not yet been included. That is a warning to revisit the design, not evidence that the measurement is adequate.

Do not treat a part tolerance as the entire permitted measurement uncertainty. Define a separate measurement-performance requirement and an acceptance rule for readings near the limit.

5. Decide which dimensions the image actually represents

A silhouette gives a projection. A through-hole may appear circular while a recessed step remains hidden behind its entrance. Likewise, a chamfer can create an edge that differs from the functional diameter specified on the drawing.

Identify the physical boundary behind every measured image edge. Ask whether it represents the entrance, exit, narrowest visible restriction, or a reflection. If that identity changes with illumination or viewing direction, the measurement needs further investigation.

Conventional perspective imaging also changes scale with object distance. In a simplified pinhole approximation, relative scale variation is approximately the distance change divided by the camera-to-object distance, provided the change is small:

This example is a geometry warning, not a lens-selection calculation. A dimensional measurement can therefore shift when a part moves vertically, even though its physical dimensions remain unchanged.

Choose machine vision lenses around working distance, field coverage, sensor coverage, and the required measurement behavior. The lens finder helps narrow those choices, but focal length alone does not establish dimensional capability.

Telecentric imaging can reduce sensitivity to object-distance changes within its specified operating range. It cannot expose an occluded boundary or turn a single projection into a complete internal measurement. Where depth matters, evaluate additional views or a suitable depth-measurement method.

6. Build a shortlist around the proven imaging geometry

Product selection should follow the visibility experiment. The parts below illustrate candidates for a stationary optical feasibility setup; they are not a validated inspection system.

For an initial silhouette trial, LS-BKL-1468 is listed as a red flat backlight. Confirm that its illuminated area covers the entire projected inspection region with presentation margin.

For an edge-formation comparison, LS-BKL-0830 is listed as a collimated backlight. Compare edge position and repeatability with your actual wall thickness and surface finish.

The relevant comparison is not simply which image looks sharper. Measure whether the detected boundary corresponds consistently to the required physical feature across accepted and rejected samples.

For stationary image acquisition, LS-CMR-0053 is a rolling-shutter GigE camera with a C-mount. Confirm the actual horizontal pixel count, exposure options, triggering behavior, and interface compatibility before using it in a measurement design.

The LS-FFL-0567 is a C-mount lens. Confirm field coverage, focus range, working distance, and usable image quality for the selected sensor and measurement plane.

Matching mounts is necessary but does not establish optical compatibility. Do not infer a complete camera-and-lens pairing from those descriptions alone.

Use the area scan cameras category and camera finder to compare acquisition options. If the part moves during exposure, evaluate motion blur and shutter behavior explicitly. A global shutter avoids row-dependent exposure timing, but it does not eliminate blur caused by excessive exposure duration.

7. Validate against independent evidence

A clear image is not an acceptance study. Compare measurements and defect decisions with an independent reference appropriate to the feature. Depending on the task, that might involve a suitable gauge, an established measurement process, or confirmed defect samples.

A useful initial study separates three sources of variation:

1. Repeated acquisition without moving the part. This reveals image noise, illumination variation, and processing instability.
2. Removal and replacement of the same part. This adds presentation and fixture variation.
3. Different parts across the intended production range. This exposes material, finish, geometry, and defect variation.

As an illustrative pilot, capture 30 images without moving one part, then repeat acquisition after 10 separate placements. These quantities can reveal obvious instability, but they are not a statistical qualification plan or proof of low escape rates.

Include defects near the acceptance boundary. Large, obvious defects alone make an inspection look easier than the production problem. Record both missed defects and acceptable parts rejected incorrectly.

Save the raw image, acquisition settings, part identity, reference result, and inspection output together. A rejected image without a traceable part and reference decision has limited value during investigation.

The lens-and-lighting evaluation plan provides a useful structure for controlled comparisons. Change one variable at a time before combining improvements.

8. Recognize when ordinary optical inspection is the wrong answer

Stop expanding an optical prototype when the required feature never produces a dependable image signal. More pixels cannot recover a boundary hidden behind an opaque wall, and image enhancement cannot establish an unobserved dimension.

Reconsider the sensing route when:

• The defect lies behind opaque material with no suitable viewing access.
• A tortuous passage hides the region that determines acceptance.
• The requirement concerns a complete volume rather than a visible surface or projection.
• Material scattering removes the contrast needed to locate an internal boundary.
• The required uncertainty remains below the demonstrated system capability.
• Inspection time cannot accommodate the views needed for adequate coverage.

A hybrid process may be appropriate. Optical inspection can check accessible openings and external datums while another method examines concealed features. Keep the decisions distinct, and validate any relationship between an external observation and an internal condition.

For some components, changing the manufacturing sequence is more effective than adding a more complex sensor. Inspecting a feature before assembly may provide direct access that disappears later. Confirm that subsequent operations cannot introduce the defect being screened.

9. Complete the ordering and acceptance checklist

Before committing to hardware, check that the proposed station has a defensible answer to each item:

• Feature definition: the rejectable condition is described physically, including location and minimum relevant size.
• Visibility: the required feature remains observable across the full presentation range.
• Coverage: all required regions are inspected, not just the easiest view.
• Sampling: field coverage and pixel count support the smallest relevant feature.
• Measurement geometry: each image edge corresponds to the intended physical boundary.
• Uncertainty: the error budget and borderline acceptance rule are defined separately from part tolerance.
• Timing: exposure, acquisition, processing, and handling fit the available inspection cycle.
• Optical compatibility: sensor coverage, mount, focus range, and working distance have been checked.
• Electrical compatibility: light power, control method, and trigger behavior have been verified from the selected documentation.
• Validation evidence: acceptable, defective, and borderline samples have independent reference decisions.

Start the camera configuration process once the imaging route is demonstrated. If visibility or measurement feasibility remains uncertain, send the feature drawing, material description, required tolerance, sample images, and inspection-time requirement through applications support before ordering.