ECG image digitization: remove grids and recover traces

A photographed ECG chart contains a waveform, a grid, labels, and camera distortion in the same image. Converting everything to grayscale first can discard the color difference that makes separation possible. This procedure is for engineering digitization of printed charts: preserve color, rectify the paper, identify trace candidates, and recover coordinates with explicit uncertainty. The output is an approximation of the printed curve, not a validated clinical recording.

Step 1: Define the output and preserve the evidence

Start by deciding what you need to recover: a visual overlay, relative pixel coordinates, or calibrated time–amplitude points. Absolute units require a documented chart scale. A waveform that looks plausible is not evidence that the conversion is accurate.

1. Preserve the original color image without additional compression.
2. Select one lead and one continuous chart segment. Exclude headings, neighboring traces, and calibration marks from the tracking region.
3. Record the printed speed, amplitude scale, and any calibration annotation separately.
4. Keep intermediate images, transformation parameters, and extraction settings alongside the exported points.

Do not infer chart speed or gain from familiar-looking grid paper. If either value is unavailable, export pixels or grid divisions and mark the physical scale as unknown.

Step 2: Rectify the chart before measuring coordinates

A rotated chart needs alignment; a tilted camera view generally needs perspective correction. Lens distortion is a separate effect. Straight grid lines that bow near the image edges suggest lens distortion or paper curvature rather than simple perspective.

Use this sequence:

1. Correct significant lens distortion using calibration appropriate to the camera, lens, focus, and image dimensions.
2. Identify well-separated grid intersections across the selected chart region.
3. Assign destination coordinates with the correct horizontal and vertical grid spacing.
4. Estimate a planar transformation and warp the color image once.
5. Check additional intersections that were not used to fit the transformation.

Four non-collinear correspondences can determine a homography, but additional points help identify poor selections. Do not map an arbitrary quadrilateral to an arbitrary rectangle: that can make the chart look square while changing its measurement scale.

The image-marker calibration procedure explains the relationship between image coordinates and known physical distances. For optical correction, the image-undistortion pipeline guide explains why calibration must match the image geometry.

Measure grid spacing near the left, center, and right of the rectified region. If spacing changes systematically, investigate the transformation or paper shape. A planar warp cannot fully flatten curled paper. The calibration-target flatness guide discusses why departures from a plane matter.

Resampling also changes thin lines. Apply the final warp to the original color image rather than repeatedly rotating and resizing intermediate images. Follow the antialiasing guidance for image warps when selecting interpolation and output resolution.

Step 3: Establish pixel-to-time and pixel-to-amplitude scales

Measure spacing over several grid intervals rather than one interval. For example, a span of 200 pixels across 10 mm gives 20 pixels/mm. A one-pixel endpoint error is then approximately 0.5% of the measured span, instead of 5% across a single 20-pixel interval.

Define horizontal and vertical scales independently after rectification:

Here, v is the documented chart speed in mm/s, and g is the documented amplitude scale in mm/mV. Image row coordinates increase downward, which explains the reversed sign in the amplitude equation.

The horizontal origin x_0 may represent a relative start time. The vertical reference y_0 must be defined explicitly. A constant reference row is not the same as estimating a drifting baseline.

Do not automatically subtract a moving average to make the waveform look centered. Baseline correction changes the recovered curve. If required for a separate analysis, retain the uncorrected coordinates and record the correction method.

Step 4: Separate the dark trace while color remains available

For a dark trace over a reddish grid, use the color difference before grayscale conversion. Work with normalized channels between 0 and 1, and examine representative patches of paper, grid, trace, and grid–trace intersections.

A useful starting point is:

These are example starting thresholds, not universal settings. Dark neutral ink generally has low values in every channel. A reddish grid commonly has a larger red channel than its green and blue channels. JPEG compression, white balance, shadows, and overlapping ink can weaken that distinction.

Prefer a graded trace score based on darkness and color neutrality over irreversible deletion. Preserve the original color pixels and let the tracker resolve uncertain intersections. An absolute red rejection rule can erase genuine trace pixels where the inks overlap or compression mixes their colors.

Check the candidate mask against the color image at quiet segments, steep slopes, peaks, and grid intersections. The live binary-preview guide describes a useful inspection pattern: display the source and thresholded view together while adjusting segmentation.

If the grid and trace are both neutral gray, color separation cannot solve the problem. Estimate periodic grid locations and reduce their influence cautiously. Removing every long horizontal or vertical structure can also remove flat waveform segments and steep deflections.

Step 5: Track a path rather than selecting one dark pixel

Choosing the darkest pixel independently in each column often selects a grid line, label, dust mark, or neighboring lead. Instead, identify candidate dark runs within each column and score their continuity across the image.

For a compact run, a darkness-weighted row centroid provides a subpixel estimate of its center. Retain multiple candidates when several disconnected runs exist. Do not average separate runs together: the resulting coordinate may lie in blank paper.

A path tracker can minimize an objective such as:

The data term favors dark, trace-like pixels. The other terms penalize abrupt displacement and changes in slope. Dynamic programming can evaluate candidate paths; a second-difference penalty requires keeping track of the preceding slope or an equivalent expanded state.

Normalize the terms before tuning their weights. Otherwise, a change in image resolution can alter the balance between pixel evidence and smoothness.

Keep transition limits wide enough for genuine steep deflections. An overly restrictive jump limit makes an attractive but incorrect rounded peak. Start with a known steep segment, then test noise rejection elsewhere. Treat these settings as image-tracking parameters, not physiological acceptance rules.

Near-vertical strokes may occupy many rows in one column. Their centroid does not necessarily represent a unique original sample. Mark these columns as geometrically ambiguous and retain the stroke extent or competing candidates when that information matters.

Step 6: Interpolate only short, defensible gaps

A missing candidate is preferable to a confidently invented waveform. Represent missing values explicitly before considering interpolation.

For a short gap bracketed by trusted endpoints, linear interpolation is transparent:

Set the maximum gap using the calibrated time scale, not only a pixel count. As an illustrative policy, permit no more than three missing columns when each column represents 2 ms. That limits the missing portion to 6 ms, but it does not prove that interpolation preserves a narrow peak.

Only fill a gap when both endpoints are credible, the intervening image contains no competing trace, and the local shape supports the connection. Do not interpolate automatically across a suspected peak, a lead crossing, a fold, or a label. Do not extrapolate beyond the first or last trusted point.

Record interpolated samples separately from directly observed samples. Spline interpolation may look smoother, but it can introduce overshoot unsupported by the printed trace.

Step 7: Work through a calibrated example

Assume a chart explicitly states a speed of 25 mm/s and an amplitude scale of 10 mm/mV. After rectification, measured spacing is 20 pixels/mm in both directions.

The horizontal conversion is:

The vertical conversion is:

A tracked point 250 pixels to the right of the time origin and 80 pixels above the reference row therefore becomes:

A vertical localization uncertainty of ±2 pixels corresponds to ±0.01 mV in this example. That is only the localization contribution; scale error, reference-row error, perspective residuals, and printing artifacts remain separate.

For acquisition planning, suppose the desired chart width is 100 mm and the target sampling density is 20 pixels/mm:

Use the dimensional-inspection resolution guide to distinguish pixel sampling from optical detail. Extra output pixels created by upsampling do not create new trace information.

Step 8: Verify the extraction and export uncertainty

Reproject the recovered points onto the rectified color image. Inspect both the complete segment and enlarged views of peaks, low-contrast regions, grid crossings, and every interpolated gap.

Also map selected points back onto the original photograph using the inverse transformation. This helps expose errors hidden by a visually convincing rectification.

Use an export table containing at least:

• x_px and y_px for the recovered image coordinates.
• time_s and amplitude_mV only when their scales are documented.
• A status such as observed, interpolated, ambiguous, or missing.
• A tracking score and an estimated localization range where available.
• References to the source image, calibration, and processing settings.

A tracking score is not a probability of clinical correctness. Calibrate any numerical confidence claim against independently labeled examples before treating it as an accuracy estimate.

Before accepting the export, manually verify several known grid displacements and independently selected trace points. Repeat the extraction with modest threshold changes. Large coordinate changes indicate fragile segmentation, even when one overlay looks acceptable.

Step 9: Improve capture before ordering hardware

For repeated digitization, hold the camera perpendicular to flat paper and use even illumination without clipped highlights. Frame the chart region with enough border to locate grid intersections. Disable changing automatic settings once a suitable capture has been established.

Match the lens field of view to the required chart width and sampling density. The machine vision lenses category and lens finder help narrow optical choices; the area scan cameras category and camera finder help compare acquisition options.

The LS-CMR-0053 is a camera candidate to evaluate for a repeatable capture station.

The LS-FFL-0602 is a lens candidate to evaluate against the required field and working distance.

These are evaluation candidates, not a verified application pairing. Confirm sensor coverage, mount compatibility, focus range, edge sharpness, and exposure performance before ordering. Nominal resolution alone does not establish extraction accuracy.

Use this final checklist:

• Preserve color until grid and trace evidence have been separated.
• Verify geometric correction against unused grid intersections.
• Confirm both physical scales from the chart documentation.
• Review steep strokes and all filled gaps manually.
• Export missing values and uncertainty instead of hiding them.
• Retain an overlay and the unchanged source image for audit.

For a capture station, start with the camera configurator. For an existing photograph, use the same checklist first; if geometry, scale, or trace visibility remains uncertain, request application support before selecting replacement hardware.