3D surface blob analysis for dents on tilted parts

A dent can be clear in height data even when its grayscale appearance barely differs from the surrounding surface. The challenge is separating that dent from part tilt, intended steps and measurement noise. This guide explains how to build a reference surface, segment signed deviations and measure connected defect regions. It is intended for inspection teams developing repeatable pass/fail decisions from calibrated 3D height maps, rather than simply highlighting unusual pixels.

Decide what counts as a surface defect

3D surface blob analysis groups neighboring measurements according to their deviation from an expected surface. Negative deviations identify depressions; positive deviations identify raised features. You then measure each connected region instead of evaluating isolated pixels.

The reference is the central decision. A fixed height threshold may work when every part sits at the same height and orientation. A tilted part usually needs a fitted plane, while a stepped part needs separate references for its intended faces.

Define the defect before selecting the threshold. Specify its polarity, minimum relevant depth, footprint and location. A harmless tooling mark and a rejectable dent can have similar peak depths but different areas.

Also distinguish localized defects from broad shape errors. If an apparently flat surface bows across most of the inspection region, use the checks in diagnosing point-cloud plane curvature. Fitting away that bow could hide a genuine form defect.

Fit the reference from sound material

Start with calibrated coordinates in consistent physical units. Pixel indices alone are insufficient when you need defect area or volume, particularly if sampling varies across the field.

Select reference regions that should remain undamaged. Distribute them across the face so the fit is constrained in both lateral directions. A narrow strip can produce an unstable plane estimate across a much wider inspection area.

For a surface represented as height over lateral coordinates, a convenient model is:

Here, r_z is the vertical residual and h is signed distance to the plane. The coefficients a and b describe slope; c describes offset. Positive h denotes a bump when positive height points toward the exposed side of the part.

An ordinary least-squares fit minimizes vertical residuals. For measurement noise that acts in several coordinate directions, an orthogonal plane fit may be more appropriate. Whichever method you choose, verify its residuals on known-good material.

Exclude invalid depth samples and known features before fitting. Robust fitting can reduce the influence of isolated outliers, but it cannot reliably identify the intended plane when most available points belong to a defect or another face.

Save the fitted coefficients and reference-region coverage with each result. They help distinguish a genuine defect from a misplaced part or a failed reference fit.

Handle tilted and stepped surfaces explicitly

Subtracting the plane removes the expected height ramp caused by tilt. However, vertical depth and normal depth are not interchangeable on a sloping face.

For example, a face tilted by 30 degrees has a vertical-to-normal conversion factor of approximately 0.866. A vertical depression of 0.20 mm corresponds to a normal depression of about 0.17 mm:

State which direction defines the drawing requirement and inspection result. A threshold specified normal to the face must not be applied directly to uncorrected vertical residuals.

Unexpected tilt across the entire acquisition may indicate a coordinate or calibration problem rather than ordinary part placement. Follow the checks for a tilted laser-triangulation point cloud before compensating for it in software.

For stepped surfaces, label each intended face and fit its reference separately. Do not force one plane through multiple terraces. Otherwise, the intentional step becomes a large false blob, and the compromised fit can reduce sensitivity elsewhere.

Create a boundary exclusion band where mixed measurements or occlusion make edge data unreliable. Set its width from observed edge behavior, not an arbitrary percentage of the image. Record the excluded area as uninspected; it is not automatically acceptable material.

A moving local reference can follow gradual shape changes, but its window must be substantially wider than the defects being sought. If its spatial scale approaches the dent width, it can follow the depression and remove the signal.

Set thresholds from repeatability and acceptance limits

Measure noise using repeated scans of representative good parts with the production acquisition settings. Include dark, reflective and steep regions that the inspection must actually evaluate.

At each location, estimate temporal variation after applying the same alignment and reference-fitting process used in production. Separately examine persistent spatial bias, which repeated averaging will not necessarily remove.

A useful starting model is:

Here, sigma_repeat is measured repeatability, k is a chosen margin and B_allowance covers residual bias or reference uncertainty. This is an engineering starting point, not a guaranteed false-alarm probability.

Suppose repeatability is 0.02 mm and the additional allowance is 0.03 mm. With a margin of 3:

For dents, mark valid points where h < -0.09 mm. For bumps, use h > 0.09 mm. Keep the two masks separate so adjacent positive and negative features do not become one ambiguous region.

A three-standard-deviation threshold still produces excursions in a large image, even under an ideal Gaussian model. Real depth errors can also be correlated or contain outliers. Connected-region size and repeatability checks therefore remain necessary.

Detection and rejection thresholds serve different purposes. Detection should capture candidate geometry; rejection should implement the approved acceptance requirement. If the noise-based detection threshold is above the smallest rejectable defect, improve acquisition or reference accuracy rather than lowering the threshold without validation.

For line-based measurements, address laser-line extraction noise before trying to remove its effects with increasingly aggressive blob filters.

Convert the threshold mask into measurable regions

Run connected-component labeling on the valid threshold mask. On a rectangular grid, four-neighbor connectivity joins only edge-sharing pixels; eight-neighbor connectivity also joins diagonal contacts. Select one convention and retain it throughout validation.

For irregular point clouds, establish a justified neighborhood distance or resample onto a calibrated grid. A point-count threshold alone is not a physical-area threshold when sampling density changes.

Measure each connected region using quantities that answer the inspection requirement:

• Footprint area in a declared coordinate plane.
• Maximum signed depth or height.
• Mean deviation over valid region samples.
• Centroid and location relative to part features.
• Length, width and orientation in the relevant surface frame.
• Integrated volume relative to the chosen reference.
• Valid-data coverage and contact with exclusion boundaries.

For uniform samples on a grid aligned with the reference plane:

Here, N is the number of selected cells and dx and dy are calibrated grid spacings. The volume estimate covers the selected region, not necessarily the full physical dent.

If you retain a grid in the original XY coordinates instead, vertical volume integration uses -r_z and XY cell area. Do not multiply normal depth by projected XY area and call the result a correctly compensated volume.

Morphological opening can remove small islands, while closing can connect nearby regions. Both alter measured geometry. Keep the original residual data, document every mask operation and verify that thin or closely spaced defects are not removed or merged.

Work through a dent-sizing example

Consider a nominally flat face inspected after alignment and resampling into a plane-based coordinate frame. Assume uniform grid spacing of 0.10 mm in both directions and the 0.09 mm detection threshold calculated above.

A connected dent region contains 240 valid cells. Its mean normal depression is 0.14 mm, and its maximum depression is 0.22 mm.

Suppose an illustrative acceptance rule rejects a dent only when its maximum depth exceeds 0.18 mm and its detected area exceeds 1.5 mm². This region exceeds both limits and would be rejected under that rule. Those example limits are not recommendations for a particular part.

The reported volume excludes shallow material outside the thresholded boundary. Lowering the segmentation threshold would generally enlarge the measured footprint and change the integrated volume. Compare results only when reference fitting, thresholding and filtering remain consistent.

Sampling also matters. A feature 1.2 mm wide spans approximately 12 sample intervals at this grid spacing:

That is a sampling calculation, not proof of optical resolution or measurement accuracy. Interpolation can produce a finer-looking grid without creating additional measured detail.

If acceptance depends on full dent volume, validate a boundary definition that includes the shallow shoulders. Do not assume the high-confidence detection mask also describes the complete defect boundary.

Choose acquisition hardware before tuning filters

The segmentation method applies to several acquisition approaches, but each must provide usable data over the required surface. Browse 3D sensors by acquisition type, then use the camera finder to narrow the hardware shortlist.

The MIS-DP3850-01P is a laser profile sensor candidate for evaluating a profile-based acquisition approach. Confirm the complete sampling, motion and measurement requirements before selecting it.

With a moving part, acquisition must preserve the relationship between lateral samples and travel position. Uneven motion or incorrect position scaling can stretch a blob and corrupt its area. Evaluate the actual scanning arrangement rather than assuming the output grid is geometrically uniform.

The MV-DPS200P-02 is a projected structured-light 3D camera candidate for evaluating an area-based acquisition approach. Confirm suitability for the surface and required acquisition conditions through application review.

Neither candidate is asserted here to include a particular surface-blob software function. Confirm data access, processing location and available measurement tools separately from the sensor's acquisition type.

Before ordering, establish the required field coverage, smallest relevant defect, depth repeatability, cycle time and acceptable missing-data fraction. Use the 3D camera working-distance guide when arranging the measurement geometry.

If dark material produces weak or incomplete depth, address stereo depth lighting and surface visibility where applicable. Software cannot recover a trustworthy dent profile from measurements that were never acquired.

Treat missing depth as unknown, not as a dent

Invalid samples need a separate mask from measured height. Replacing missing depth with zero can create an artificial depression; replacing it with the fitted plane can hide a real defect.

Track missing-data coverage for the inspection region and for each candidate's neighborhood. If a potential dent intersects a large invalid patch, report an incomplete measurement or request another acquisition according to the inspection policy.

The guide to missing points in a 3D point cloud helps distinguish acquisition gaps from processing errors.

Small-hole interpolation may be acceptable for visualization, but measurement use requires validation against the smallest relevant defect. Keep the original validity mask even when interpolation is applied.

Edge-adjacent dents are particularly difficult. Occlusion can remove the deepest area, while mixed edge measurements can create extreme values. Neither an apparently shallow result nor a single deep sample is sufficient evidence without checking coverage.

Verify performance across realistic variation

Build a validation set that includes sound parts, known defects, allowed features and difficult surface conditions. Cover position and tilt variation as well as defects near steps and inspection boundaries.

Repeated scans of one stationary sample measure only part of the problem. Remove and replace parts to expose fixture and reference-fit sensitivity. Include the movement and acquisition timing expected during operation.

Record both missed defects and false rejects. A filter that removes every noisy island may also remove narrow dents; a reference that produces tidy residuals may be following broad damage.

Retain raw data and settings so failed decisions can be reproduced. Follow the guidance on recording depth data losslessly when building an offline validation set.

Check the following before releasing the inspection:

• The residual sign agrees with a known dent and a known bump.
• Coordinates and spacing use consistent physical units.
• Reference regions remain valid across permitted part movement.
• Intended steps are labeled or excluded explicitly.
• Missing depth cannot silently become acceptable material.
• Connectivity and filtering preserve the smallest required defect.
• Area and volume use the declared coordinate convention.
• Acceptance decisions remain stable across repeated placements.

Know when plane-based analysis is the wrong answer

A plane is not an adequate nominal model for a deliberately curved face. Use a suitable geometric model or registered reference surface, then evaluate deviations in a clearly defined direction.

Likewise, surface-blob thresholds are not automatically a surface-roughness measurement. Roughness requires its own filtering, evaluation length and parameter definitions. A method designed to identify localized dents can suppress or misinterpret fine texture.

If defects occupy most of the face, there may be too little sound material for an independent plane fit. If missing depth covers the critical region, a reliable sizing result may be impossible from that acquisition. In both cases, report the limitation rather than forcing a numeric answer.

For a hardware review, bring a representative height map, valid-data mask, defect dimensions and acceptance rule to application support. Start the camera configurator once those requirements are defined, so acquisition capability and the measurement method are selected together.