SurView DIC

Software for digital image correlation: full-field displacement and strain measurement from speckle images.

Free and open source, LGPL-2.1-or-later · Qt and VTK · correlation by OpenCorr

The SurView DIC workspace: a speckled tension specimen with a measured displacement field over it, the image record on the left, correlation settings on the right, and a per-point readout beneath them.

Capabilities

SurView DIC measures how a speckled specimen moved and deformed between photographs, over a whole loading sequence, reads it back as a loading curve, and reports how far each measured point can be trusted.

Two-dimensional subset correlation over a regular grid of points
Region of interest drawn by hand or segmented automatically from speckle quality
Whole loading sequences measured against one reference, in frame order
Reference updating when correlation against the original frame degrades
A second pass that repairs points the first solve could not measure, and says which they were
Strain fitted from neighbouring displacements, in three measures
Per-point noise floor and match conditioning, computed on every run
Per-point readout of every measured channel, pinnable
Any quantity plotted against frame, over the whole sequence
Virtual extensometers placed with two clicks, read as a loading curve
Provenance recorded for every image, verified by SHA-256
Export to VTK unstructured grid and to comma-separated text
One numbered file per frame, which ParaView groups as a time series

Specification

What the correlation actually does, in the terms the literature uses.

Solvers

Inverse-compositional Gauss-Newton (ICGN), Newton-Raphson, inverse-compositional Levenberg-Marquardt (IC-LM). Every offered combination is covered by a test.

Shape functions

First order (affine) and second order (quadratic), where the solver implements it.

Interpolation

Bicubic B-spline, over the whole dense image.

Subset and grid

Subset radius and grid step chosen per run; the grid is held one subset radius clear of every image edge.

Strain

Fitted by plane regression over a subregion of chosen radius. Cauchy (small strain), Green-Lagrange and Euler-Almansi. Points correlating below 0.9 are excluded from every fit.

Repairing failures

Points the first solve rejected or correlated poorly are re-attempted from a displacement fitted to their reliable neighbours, then solved again in full, in rounds. An answer is kept only if it is better than the one it replaces, and every repaired point is marked wherever it is reported. On the example data it raises a hard frame from 5401 measured points to 8455.

Plots

Any field quantity against frame, and virtual extensometers: two points on the specimen, reported as gauge length, elongation and engineering strain. A frame that cannot be read breaks the curve rather than being joined across.

Reliability

Displacement noise floor (DIC's sigma, in pixels) and match conditioning (DIC's beta, dimensionless), for every point of every run.

Reads

TIFF, PNG, JPEG and BMP, 8-bit and 16-bit, greyscale. Row order is normalised to the file's own; nothing is converted or rescaled.

Writes

.vtu VTK unstructured grid, and .csv. Both carry both images with their SHA-256, the solver and its settings, the region, the strain fit and the engine revision.

Coordinate frame

Image pixels throughout, x right and y down, origin at the top-left pixel centre. The viewport, the region, the engine and the exported file all use it, and the file states it.

Engine

katalystnord/OpenCorr, branch surview-dev: our working fork of OpenCorr (Jiang, MPL-2.0). Required, and not interchangeable with upstream today - auto-detected regions, per-point reliability and per-point failure reasons are all capability work that lives in the fork. Pinned to an exact commit, recorded in the repository and compiled into every export.

Built with

Qt 6, VTK 9.5, OpenCV 4, Eigen 3.4, FFTW 3, nanoflann, OpenMP. C++20.

The record, and the measurement

A measurement is worth what the record behind it is worth, so SurView keeps both in view.

What was recorded

Every image is recorded as it arrived: path, size, modification time, SHA-256, the decoder that read it, its pixel type, and how much of that type's range the data actually used. Nothing is converted, resampled or normalised on the way in.

What was measured

Nothing unmeasured is reported as a number. A rejected point, an unfitted strain, an unestablished uncertainty and a frame an extensometer could not read are absent or not-a-number, never zero, in the display, in every plot and in every exported file. Failures are counted by the engine's own reason, and a point repaired by the second pass is marked as one everywhere it appears.

Screenshots

From a running build, measuring the example data that ships with it.

A five-frame tension sequence in SurView DIC: the displacement field over the speckled specimen with a virtual extensometer drawn across it, and beneath it a plot of engineering strain against frame rising smoothly from zero to 0.055.

A loading curve, from two clicks

Place a virtual extensometer on the specimen and SurView reads the change in the distance between its two points across every frame, as gauge length, elongation or engineering strain. It measures the distance BETWEEN the anchors, so a specimen that simply moved reports no strain at all. A frame the gauge cannot read breaks the curve and is counted in words beneath it, rather than being joined across as though it had been measured.

This is the synthetic tension set, whose exact strain per frame is stated in the data that ships with it. Over the first four frames the gauge reads 0.005, 0.010, 0.020 and 0.036 against stated strains of 0.005, 0.010, 0.020 and 0.035, and a test in the repository holds it to that.

The repair map on a real dogbone specimen: most of the field is slate blue for points measured on the first solve, with orange patches around the two holes, along the edges and in a band across the gauge length where the second pass repaired them. The colour bar has exactly two blocks, labelled 0 and 1.

Which points were hard won

Most points a solve loses were given a poor starting guess rather than being unmeasurable. A second pass fits the displacement of the reliable points around each failure and solves it again from there, in rounds, so repair spreads inward from ground that correlated well. Here it recovered 2531 of 9044 measured points.

A repaired point is a real measurement with its own correlation, and it is marked as one on the map, in the readout, in the run report and in both export formats. Where the repairs fall tells you something the count cannot: clustered at the holes and edges, as they are here, is the specimen talking.

ParaView with the exported .vtu open: the pipeline browser shows gauge.vtu, the properties panel lists displacement, strain, correlation, solved and both reliability arrays, and the strain field is drawn on the specimen with unmeasured points left as visible holes.

The exported field, in ParaView

SurView's own export, written from the File menu and opened with no conversion step: the format is VTK's own. Every channel the run measured is in the array list, and so are the provenance strings.

The gaps are unmeasured points. They are not-a-number rather than zero, so they arrive as holes rather than as a flat region no reader could tell from a measurement.

A rigid rotation measured as strain: the map is near-uniform white on a diverging scale, running from -0.000457 to 0.000116.

A rotation is not a strain

A specimen turned, not deformed. Strain runs from -0.000457 to 0.000116 across the whole field: zero, to within the method's own noise.

The scale is centred on zero because strain's zero is a physical state. Displacement's is not, so it is scaled differently.

The point readout panel showing reference pixel, displacement, magnitude, correlation, noise floor sigma, match conditioning beta and fitted strain, each with a sentence saying what it cannot tell you.

Every number, and what it is not

Point at the field for that point's own measurement: displacement, correlation, the noise floor of its subset, how well conditioned the match was, and the strain fitted around it. Click to pin a reading.

Each value carries one sentence on what it does not cover. The noise floor is a lower bound on error, not a total error bar, and it never examines the target image; the conditioning is comparable within a run and meaningless across runs.

Region drawing mode: a bounded quadrilateral over the speckle, with an on-screen bar reading four corners placed, the pointer position in image pixels, and buttons for undo, close and cancel.

Regions, drawn or detected

Click corners to bound the specimen, or let SurView segment the speckled area from an image-gradient quality map. The mode carries its own instructions, a live corner count, the pointer's position in image pixels, and every way to finish it.

Example data

Two kinds, because they answer different questions. Both ship with the source.

Six synthetic sets, with the exact answer

Rendered from an analytic speckle pattern rather than warped from an image, so no pixel is ever resampled and the displacement is exact by construction. Each frame carries its deformation gradient, its displacement at the corners and centre, and a noise floor measured from its own pixels.

Sub-pixel translation, 0.25 to 2.5 px
Uniaxial tension with Poisson contraction
Rigid rotation to 15 degrees, no strain at any step
Simple shear
Large strain to 30 per cent, for reference updating
Movement straddling what the speckle can resolve

Three real experiments

Photographs of actual specimens, with what a real camera brings: lighting, grips, out-of-plane motion, and background carrying no speckle at all.

Aluminium dog-bone in uniaxial tension
Perforated specimen, stress concentration at the holes
Speckled plate in rigid-body rotation

From the pyALDIC project, BSD-3-Clause, cited with the data.

Download

SurView builds from source against a pinned engine revision, so a build is always traceable to one exact correlation library. Both clones are required: the second is the correlation engine, and SurView will not configure without it.

git clone https://github.com/katalystnord/SurView.git
git clone -b surview-dev https://github.com/katalystnord/OpenCorr.git

cd SurView
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build

./build/src/SurView examples/real/01_tension_without_holes/image_000*.png

The last line opens a real tension sequence with its reference and four loaded frames already imported. Press Run Correlation.

The engine clone is katalystnord/OpenCorr rather than upstream on purpose. It is our working fork of Jiang's OpenCorr, and several things SurView shows on screen are implemented there. Upstream will not build this application as it stands. Fixes that are generally useful go upstream as pull requests rather than staying in the fork, and a script in the repository checks all three links - upstream to fork, our upstreamed fixes by content, and fork to the pinned commit.

Sister tool

SurView DIC and PatternFab are two ends of one workflow: the speckle pattern PatternFab makes is the pattern SurView measures.

Measure the pattern

SurView DIC

Speckle images in, displacement and strain fields out, with every point's reliability reported and full provenance carried into ParaView and FreeCAD.

This tool.

Fabricate the pattern

PatternFab

An optimized speckle pattern into a 3D-printed stamp, a laser-cut stencil, a decal or a direct-write file, with the manufacturing constraints applied before anything is made.

katalystnord.github.io/patternfab

Licence and credits

Correlation is by OpenCorr (Jiang, MPL-2.0): 2D, stereo and volumetric DVC in one library, and the work to cite for the method.

SurView builds against our working fork of it, which carries the capability work this application needs. It is a staging area for that work and for fixes on their way upstream, not an attempt to run a competing version of OpenCorr.

The application shell is Qt and everything rendered or exported goes through VTK, which is why results open natively in ParaView and FreeCAD rather than through a converter.

SurView DIC is LGPL-2.1-or-later, the licence FreeCAD chose in the same ecosystem. No paid tier and no open core.