Through-hole solder inspection beyond golden-image matching

Through-hole solder joints can look different without being defective. Lead position, solder volume, board finish, and reflections all change the image, making a single reference photograph a poor acceptance boundary. For teams moving from manual inspection to automated checks, the useful starting point is a defect list with measurable evidence. You can then select views, lighting, and thresholds that distinguish unacceptable conditions from normal variation, while keeping hidden joint conditions outside the claims of visible-light inspection.

Define the inspection decision before choosing hardware

Use defect-specific measurements rather than a single similarity score. A reference image remains useful for locating pads and leads, but its exact brightness pattern should not define acceptable soldering.

Start with the approved assembly requirements and examples approved by your quality team. Record which conditions are acceptable, which require rejection, and which need additional inspection. Do not substitute camera thresholds for the applicable workmanship requirements.

A decision record should contain the defect name, region to inspect, measurement, units, acceptance boundary, and handling of uncertain cases. Different joint families may need different records.

For fixture and viewing considerations, use the related PCB inspection setup guide. The emphasis here is deciding what the image must prove, rather than matching its overall appearance.

Separate visible defects from hidden joint conditions

A camera sees exposed surfaces along its viewing direction. It cannot establish internal barrel fill, internal void content, or the integrity of a concealed interface from an underside photograph alone. Electrical continuity testing answers another question and does not establish complete visual or structural acceptability.

Treat a visible surface pit separately from an internal void. A dark patch may instead be a reflected feature, flux residue, a shadow, or an exposed opening. Change the illumination direction and inspect whether a stable geometric boundary remains.

A perpendicular view is often useful for bridges, particles, pad registration, and projected fillet boundaries. Oblique views can reveal features hidden behind a lead, but introduce their own shadows and perspective effects. Add them only when they provide evidence needed for a defined defect.

Where an internal condition determines acceptance, evaluate an appropriate complementary method, such as X-ray inspection or destructive examination of process samples. Its suitability depends on the joint geometry and the actual requirement.

Size the image from the smallest required feature

Choose area scan cameras by the field of view and smallest feature that must be distinguished. Use the camera finder after calculating the sampling requirement, not before.

As a starting estimate, allocate several pixels across a small feature. Five pixels across its width can be a useful feasibility target, but it is not an acceptance standard or a guarantee of detection. Contrast, lens performance, focus, noise, and segmentation still determine usable information.

Consider a hypothetical inspection field measuring 80 mm across and 60 mm vertically. Suppose the smallest relevant surface opening is 0.10 mm wide, and the feasibility target is 5 pixels across it.

This is a sampling estimate, not proof that every opening of that size will be detected. Verify representative features near the image corners and beside reflective leads, where performance may differ from the center.

The LS-CMR-0053 is a rolling-shutter camera candidate for a stationary-board evaluation. Confirm acquisition timing and illumination behavior before treating it as a suitable production choice.

For moving boards, evaluate shutter behavior, exposure time, and motion blur separately. A stationary prototype does not establish moving-line performance.

Choose optics around coverage, focus, and access

Select machine vision lenses for sensor coverage, working distance, field of view, and the height range of visible features. The lens finder helps narrow candidates, but mechanical access around the solder side must also be checked.

Do not calculate physical sensor width directly from a nominal optical-format label. Obtain the active sensor dimensions when estimating focal length, then verify the actual field with the intended lens and mounting arrangement.

The LS-FFL-0567 is an optics candidate to assess against those requirements. Its listing alone does not establish the field of view or focus range for your station.

Lead tips, board surfaces, and fillet contours occupy different heights. Test all required heights before choosing the aperture. Stopping down may improve depth of field, but reduces received light and eventually sacrifices fine detail through diffraction.

Use a stable fixture that supports the board without obstructing the inspected regions. Check the largest and smallest assemblies, not just the first sample that fits. A lead or fixture feature that hides the critical boundary makes that measurement unavailable.

The object-space sampling explanation helps connect optical magnification to feature coverage. For focus verification, the sharpness measurement guide explains why a repeatable metric is preferable to judging a shiny fillet by eye.

Make lighting reveal defects rather than random highlights

Solder is reflective, so changing the illumination direction changes the visible brightness pattern. A bright region may identify a favorable reflection angle rather than additional material. A dark region may simply reflect a dark part of the surroundings.

Start with independently controlled bar lights, positioned so each direction can be tested separately. Use the lighting finder to compare candidate geometries after recording the available mounting space.

The LS-BAR-1047 provides a standard bar-light candidate, while the LS-BAR-1276 provides a diffused candidate. Evaluate how each reveals boundaries and suppresses misleading reflections on actual joints.

The surface-versus-silhouette lighting guide explains the underlying geometry. For solder-side surface inspection, select the arrangement by defect contrast rather than by overall image attractiveness.

Lock exposure and gain during each comparison. Examine saturated pixels as well as the average brightness. Saturation removes information; a brighter image is not automatically a more useful image.

If sequential views are needed, associate them with the same board and joint. The cross-exposure coordinate matching guide addresses that association problem. Directional images can also support shape analysis, but the photometric stereo guide explains why this requires more than subtracting two pictures.

Measure registered features, not whole-image differences

Locate the board using fiducials or stable board features before evaluating individual joints. Then refine each region using the pad, lead, or another reliable local feature. Registration should tolerate normal placement variation without allowing a defect to move outside its inspection region.

Useful measurements can include:

• Minimum apparent separation between conductors that must remain isolated.
• Area and position of a candidate surface opening within the fillet region.
• Projected fillet coverage relative to the pad and lead geometry.
• Count and size of isolated particle candidates within defined exclusion zones.
• Lead displacement relative to the expected pad center.

Each measurement needs a definition. For example, specify whether fillet coverage excludes the lead silhouette, exposed pad, flux residue, and occluded regions. Otherwise, the same nominal measurement can describe different image content across recipes.

Normalize distances and areas to calibrated units where appropriate. A planar calibration supports measurements near the board plane, but elevated fillet surfaces and lead tips can have different apparent scales under perspective imaging.

Do not average away a critical bridge indication across several lighting directions. Instead, define how the individual observations combine: confirmation in another view, additional review, or direct rejection after validated classification.

If subtraction remains part of the method, first address registration and edge sensitivity. The image-subtraction artifact guide explains why displaced contours can produce apparent defects even on acceptable parts.

Set thresholds with an uncertainty band

Use approved examples to establish measurement distributions for acceptable and defective joints. Include examples near the acceptance boundary, not just obvious failures. If the distributions overlap substantially, improve the imaging or measurement before forcing a binary threshold.

Suppose an approved product requirement establishes a minimum visible conductor separation of 0.20 mm. This is an illustrative requirement, not a universal soldering rule. At the earlier sampling estimate:

A possible evaluation policy would pass measurements at or above 0.24 mm, flag measurements below 0.16 mm as likely failures, and send the intervening range for review. Those boundaries require validation against the actual requirement and measurement error.

The illustrative margin is not a measured uncertainty budget. Determine a defensible margin through repeated captures, board reseating, focus variation, and comparison with an independent reference measurement. Pixel sampling alone does not establish measurement accuracy.

A model trained from images can support candidate detection or classification. However, its training set should represent acceptable variation, and the final decision still needs validation against the defect definitions. Learning one acceptable appearance merely recreates the golden-image problem in another form.

Validate across boards, lots, and nuisance variation

Separate recipe development from final validation. Keep complete boards or production lots together when dividing data, so nearly identical crops from one board do not appear in both groups.

Challenge the inspection with lead-position variation, board bow, normal solder appearance changes, permitted residue, and repeated fixture loading. Include all relevant pad and lead geometries. An algorithm that works on one joint family may not transfer to another.

Record outcomes by defect type and joint family:

• Defective joints incorrectly accepted.
• Acceptable joints incorrectly rejected.
• Joints routed for review or declared uninspectable.
• Capture or registration failures.
• Processing time and sustained inspection throughput.

Counts need denominators. Missing none of 20 defective examples gives much weaker evidence than missing none of 1,000 representative defective examples. As a rough statistical guide, zero events in 1,000 independent trials gives an approximate one-sided 95% upper event-rate bound of 3/1,000, or 0.3%. It does not prove zero risk or cover defect types absent from the trial.

Measure nuisance variation separately from defect sensitivity. If reseating the same board changes its classification, fix registration, illumination, or mechanical repeatability before collecting a larger training set.

Document manual-review outcomes and disagreements. Review should refine the defect definitions and reveal missing evidence, rather than silently override inconvenient classifications.

Check ordering readiness and recognize the limits

Before committing the station design, complete this checklist:

• Confirm which surface and internal conditions the inspection must distinguish.
• Obtain approved acceptable, defective, and borderline samples for each joint family.
• Verify sampling and sharpness across the entire field and required height range.
• Check every critical region for lead, component, and fixture occlusion.
• Record lighting positions, exposure settings, and the image sequence.
• Validate thresholds on boards excluded from recipe development.
• Define review handling, traceability, and behavior after a capture failure.
• Confirm the full cycle time, including motion, settling, capture, processing, and review routing.

The usual single-camera answer is wrong when the decisive feature is hidden, when required surface detail cannot be resolved across the field, or when acceptable and defective conditions remain visually indistinguishable. More elaborate image processing cannot recover evidence the setup never captured.

Prepare a sample package containing board dimensions, joint photographs, the smallest required feature, approved defect definitions, and cycle-time targets. Use the camera configuration step to assemble an initial imaging selection, or contact applications support to review the sample package before ordering.