Rectified Data (Z-Map): Normalized Height and Surface Representation for Precise Analysis

The rectification pipeline consists of 3 sequential operations applied to every acquired frame. First, geometric undistortion removes lens-induced deformation from the raw image. Second, metric scaling converts pixel-unit disparity or phase values into physical height values expressed in micrometers or millimeters. Third, resampling maps the corrected data onto the target XY raster at uniform lateral spacing.

Intrinsic calibration parameters define the sensor’s internal optical geometry: focal length in x and y directions, the principal point coordinates (optical axis intercept on the sensor array), and radial and tangential distortion coefficients. Extrinsic calibration parameters define the sensor’s pose relative to a reference coordinate system: a 3×3 rotation matrix and a 3×1 translation vector. Both parameter sets are determined during sensor calibration using a calibration artifact with known geometry — typically a precision glass plate with etched dot patterns or a stepped height standard traceable to national length standards.

Applying these parameters to each acquired frame produces a corrected Z-Map in which each pixel maps to a known physical location with sub-micrometer positional accuracy, depending on sensor resolution and calibration quality.

Calibration quality determines the accuracy of every rectified output the sensor produces. A sensor with a 0.1-pixel residual distortion error after calibration introduces a systematic Z-error proportional to the sensor’s height-per-pixel scaling factor. For a sensor with 1 µm/pixel Z-resolution, that residual produces a systematic Z-offset of approximately 0.1 µm across the measurement field. Thermal drift, mechanical vibration, and aging of optical components cause calibration parameters to shift over time, requiring periodic recalibration.

Two calibration strategies apply to industrial 3D sensors: factory calibration and field recalibration. Factory calibration is performed under controlled environmental conditions at the point of manufacture and provides the highest achievable parameter accuracy. Field recalibration allows the user to update calibration parameters in situ using a portable calibration artifact, compensating for installation-specific distortions and environmental drift.

Measurement System Analysis (MSA) methods — including gauge repeatability and reproducibility studies — quantify the combined effect of calibration residuals and measurement noise on the sensor’s output. Accuracy specifications derived from MSA results define the operational limits within which rectified Z-Map data is metrologically valid.

A Z-Map stores height values as a 2D array of W × H cells, where W is the number of columns (X-direction) and H is the number of rows (Y-direction). Each cell contains one height value Z expressed in physical units, and optionally one intensity value and one confidence score. The intensity channel encodes the reflected light amplitude at each surface point. The confidence channel flags unreliable height values caused by occlusion, low surface reflectance, or sensor saturation.

4 parameters define the spatial geometry of a Z-Map: lateral pixel pitch in X (µm/pixel), lateral pixel pitch in Y (µm/pixel), Z-resolution (µm/count), and the coordinate origin relative to the sensor or world reference frame. Lateral pixel pitch is determined by the optical magnification and the sensor array pixel size. Z-resolution is determined by the sensor’s measurement principle and encoding bit depth. A 16-bit Z-Map with a 100 mm measurement range provides a raw Z-resolution of:

Zres=100mm65,535counts≈1.5μmcountZres​=65,535counts100mm​≈1.5countμm​

Z-Maps are referenced to one of 3 coordinate systems: the sensor coordinate system (origin at the sensor’s principal point, Z-axis aligned with the optical axis), the machine coordinate system (origin at a machine reference point, axes aligned with machine kinematics), or the workpiece coordinate system (origin and axes defined by workpiece datum features as specified by GD&T standards). The coordinate transformation between these frames is part of the rectification and registration pipeline.

Standard storage formats for Z-Maps in industrial metrology include 16-bit grayscale TIFF (height encoded as unsigned integer with a defined scale factor), OpenEXR (height as 32-bit floating-point), and HDF5 (hierarchical container supporting multiple channels and metadata). Proprietary sensor APIs also define vendor-specific binary formats with embedded calibration metadata.

Three terms describe raster-based height datasets in industrial imaging and computer vision: Z-Map, range image, and depth map. The terms are used interchangeably in some contexts but carry distinct meanings in precision metrology.

A Z-Map stores height values orthogonally projected onto a flat reference plane. Each Z-value represents the perpendicular distance from the reference plane to the surface, measured along the Z-axis. This orthographic projection is the correct representation for surface roughness measurement, flatness deviation analysis, and step-height measurement according to ISO 25178.

A range image stores the distance from the sensor origin to each surface point, measured along the ray direction from the sensor to that point. Range images are referenced to the sensor’s perspective and require deprojection before use in Cartesian metrology. They are the native output format of time-of-flight cameras and some 3D laser scanners.

A depth map is a range image referenced to a camera perspective, used in computer vision, robotics, and RGB-D camera systems. Depth maps represent scene geometry from the camera’s viewpoint, not from a flat measurement reference. Converting a depth map to a Z-Map requires knowledge of the camera’s intrinsic parameters and a defined measurement reference plane.

A point cloud provides a richer 3D representation in which each point carries independent X, Y, and Z coordinates without constraint to a regular grid, and is addressed in its own article within Node: Metrology.

Geometric distortion correction removes lens-induced deformation from the raw sensor image before any metric interpretation. Two distortion types affect industrial optical sensors: radial distortion and tangential distortion.

Radial distortion causes straight lines in the physical scene to appear curved in the sensor image, with barrel distortion displacing image points outward from the optical axis and pincushion distortion displacing them inward. The Brown-Conrady distortion model quantifies radial distortion with 3 to 6 polynomial coefficients (k1,k2,k3,k4,k5,k6k1​,k2​,k3​,k4​,k5​,k6​), with higher-order terms becoming relevant for wide-angle optics and short working distances. In precision industrial optics, radial distortion coefficients typically range from 10−310−3 to 10−510−5 — small values in absolute terms but producing micrometer-level Z-errors at measurement field edges if not corrected.

Tangential distortion results from lens element tilt and decentering relative to the optical axis. It displaces image points in a direction perpendicular to the radial direction and is modeled with 2 additional coefficients (p1,p2p1​,p2​). Tangential distortion magnitudes in precision industrial lenses are typically one order of magnitude smaller than radial distortion.

The undistortion operation computes, for each pixel position in the distorted image, the corresponding position in the ideal undistorted image using the inverse distortion model. Bilinear interpolation fills the corrected image at sub-pixel accuracy. After undistortion, all image points satisfy the pinhole camera model: straight lines in the scene project as straight lines in the corrected image.

Resampling maps the geometrically corrected sensor data onto the target Z-Map grid at uniform lateral spacing. The source data points, after undistortion, do not fall on a regular grid — their positions depend on the original pixel positions and the distortion correction mapping. Resampling assigns a Z-value to each target grid position by interpolating between neighboring corrected data points.

3 interpolation methods apply to Z-Map resampling, each with distinct accuracy and computational cost. Nearest-neighbor interpolation assigns each target cell the Z-value of the closest corrected data point — fast but introducing quantization artifacts at the spatial period of the source grid. Bilinear interpolation computes the weighted average of the 4 nearest corrected data points — adequate for most industrial applications with measurement uncertainties above 1 µm. Bicubic interpolation uses a 4×4 neighborhood for a smoother approximation that preserves surface curvature more accurately — preferred for roughness and waviness analysis according to ISO 25178 filter chain requirements.

Confidence masking assigns a validity flag to each target cell during resampling. Cells that fall outside the measurement range, correspond to occluded surface areas, or receive contributions from fewer than 2 valid source points are flagged as invalid (NaN or a defined fill value). Downstream algorithms read the confidence channel and exclude invalid cells from surface fitting, defect detection, and dimensional evaluation. Confidence mask coverage above 95% of the measurement field is a standard requirement for inline 100% inspection applications.

Surface reflectance variation across a workpiece affects the signal quality of the raw sensor data that enters the rectification pipeline. Surfaces with both highly specular regions (mirror-like metal faces) and highly diffuse regions (matte coatings, laser-engraved textures) exceed the dynamic range of a single sensor exposure, producing saturated or underexposed areas in the raw image.

HDR acquisition and Multiple Slope techniques address this by combining measurements from multiple exposures or multiple laser power settings into a single raw frame before rectification. The merged raw frame provides valid signal across the full reflectance range of the workpiece, increasing effective confidence mask coverage after rectification. HDR acquisition strategies and Multiple Slope laser control are addressed in the Lasertriangulation cluster articles.

Z-Map noise comprises 2 statistically distinct components: random noise and systematic error. Random noise produces measurement-to-measurement Z-value variation at a fixed surface point, following a normal distribution with zero mean and a standard deviation σZσZ​ that characterizes the sensor’s height repeatability. Typical σZσZ​ values for industrial laser triangulation sensors range from 0.1 µm to 5 µm, depending on sensor model, measurement range, and surface material.

Systematic errors produce repeatable Z-offsets that do not average out over repeated measurements. 4 systematic error sources affect rectified Z-Maps in industrial installations:

• Calibration residuals — the remaining difference between the mathematical distortion model and the actual lens behavior — produce a spatially varying Z-offset that depends on field position.
• Thermal drift shifts intrinsic calibration parameters as sensor temperature changes during warm-up or across the production environment’s thermal cycle, producing a drift in the Z-scale factor and Z-offset over time.
• Fixed-pattern noise in the camera sensor introduces a spatially periodic Z-modulation at the pixel pitch frequency.
• Mechanical vibration during acquisition introduces Z-errors proportional to the vibration amplitude at the sensor’s measurement wavelength.

The combined systematic Z-error of a calibrated sensor is quantified as the sensor’s accuracy specification, expressed as the maximum permissible error (MPE) for a reference measurement defined in the sensor’s calibration certificate. Measurement uncertainty propagation from calibration residuals to Z-Map values follows the Guide to the Expression of Uncertainty in Measurement (GUM).

Repeatability of a rectified Z-Map is the variation in Z-values across repeated measurements of the same surface point under unchanged conditions: same sensor, same part position, same temperature, consecutive acquisitions. Reproducibility is the variation in Z-values when conditions change: different sensor units, different operators, different time points, or different positions within the measurement field.

Gauge Repeatability and Reproducibility (GR&R) studies quantify both components according to AIAG MSA Reference Manual procedures. A standard 10-part, 3-operator, 2-replication GR&R study on a Z-Map sensor evaluates the proportion of total process variation attributable to the measurement system.

Linearity quantifies how Z-measurement accuracy varies as a function of Z-position across the sensor’s measurement range. Resolution defines the smallest detectable Z-difference, determined by the sensor’s signal-to-noise ratio and the Z-encoding bit depth. Accuracy defines the closeness of the measured Z-value to the true height of the surface reference. All 3 properties are addressed in their dedicated articles within Node: Metrology.

Rectified Z-Maps are transferred between the sensor, processing host, and evaluation software through 3 interface types: file-based transfer, shared memory, and streaming protocol.

File-based transfer stores the Z-Map as a TIFF, OpenEXR, or HDF5 file on local storage or a network share. 16-bit grayscale TIFF is the most widely supported format across industrial metrology software packages, including surface analysis tools compliant with ISO 25178. The Z-scale factor and lateral pixel pitch are stored as TIFF metadata tags or in a sidecar XML file. OpenEXR provides 32-bit floating-point precision and native multi-channel support (Z, intensity, confidence) in a single file without requiring sidecar files. HDF5 supports hierarchical data organization, allowing a single file to contain Z-Map data, calibration parameters, acquisition timestamps, and structured measurement results.

Shared memory transfer provides zero-copy Z-Map access between sensor driver and processing application on the same host, reducing transfer latency to below 1 ms for Z-Maps up to 5 megapixels. Streaming protocols deliver Z-Map data from sensor to processing host over Ethernet at frame rates up to 3,000 profiles per second in profile-mode sensors. GigE Vision and GenICam define the transport layer and device interface standards for streaming Z-Map data in industrial sensor networks, and are addressed in the IoT Protocols cluster articles.

Real-time rectification applies calibration parameters to each acquired frame within the sensor’s cycle time, delivering a rectified Z-Map to the host without post-processing delay. Cycle time requirements in inline 100% inspection applications range from 1 ms to 50 ms per frame, depending on conveyor speed and part size — equivalent to frame rates between 20 Hz and 1,000 Hz.

Embedded rectification on FPGA or DSP hardware within the sensor head performs undistortion, metric scaling, and resampling in a pipelined architecture that processes each pixel as it is read from the sensor array. This approach achieves rectification latency below 1 ms independent of Z-Map resolution. Host-side GPU rectification applies calibration corrections in parallel across the Z-Map using CUDA or OpenCL kernels, achieving throughput above 500 megapixels per second on current-generation industrial GPUs — sufficient for Z-Maps up to 10 megapixels at 50 Hz.

The choice between embedded and host-side rectification depends on 3 factors: available compute resources at the sensor head, required output latency, and integration architecture of the production line control system. Inline 100% inspection applications with latency requirements below 5 ms use embedded rectification. Applications with post-acquisition processing workflows — including 3D stitching, multi-sensor fusion, and batch analysis — use host-side rectification to leverage centralized compute infrastructure. Inline quality inspection system architecture is addressed in its dedicated article within this documentation.