SubsetDIC

SubsetDIC is a Python package for 2D subset-based digital image correlation (DIC). The performance-critical algorithms are implemented in C++ and exposed to Python through pybind11.

This project is based on the algorithm flow of Ncorr and keeps the original Ncorr reference source under reference_code/ for comparison.

For the pure Python implementation, see lbd-hfut/Subset-DIC.

Features

• Subset-based 2D DIC for grayscale image pairs.
• C++ core with Python API.
• Quintic B-spline preprocessing for image interpolation and reference-image gradients.
• ROI-based region extraction.
• SIFT or manual seed selection.
• Region-growing DIC propagation.
• Optional displacement-gradient / strain-field calculation.
• Example ring and star cases under case/.

Project Layout

SubsetDIC/
|-- src/
|   |-- subsetdic/        # Python API and configuration helpers
|   `-- cpp/              # C++ DIC core and pybind11 bindings
|-- config/
|   `-- default.yaml      # Default DIC parameters
|-- case/
|   |-- ring/             # Example ring images and ROI
|   `-- star/             # Example star displacement images and ROI
|-- tests/
|   |-- test_dic.py       # Synthetic translation integration test
|   `-- run_cases.py      # Run bundled ring/star cases
`-- reference_code/       # Ncorr MATLAB/C++ reference implementation

Installation

Requirements:

• Python 3.9+
• CMake 3.21+
• A C++17 compiler
• Visual Studio C++ Build Tools on Windows

Install in editable mode:

python -m pip install -e .

This compiles the C++ extension module _core. In editable mode, Python source files are loaded from this repository, while the compiled extension may be installed into your Python environment's site-packages, for example:

C:\Users\<user>\miniconda3\Lib\site-packages\subsetdic\_core.cp312-win_amd64.pyd

After changing files in src/cpp/, run the install command again to rebuild the extension.

Quick Start

import numpy as np
from PIL import Image
from subsetdic import SubsetDIC

ref = np.array(Image.open("case/star/001.bmp")).astype(np.float64)
cur = np.array(Image.open("case/star/002.bmp")).astype(np.float64)
roi = (np.array(Image.open("case/star/003.bmp")) > 128).astype(np.uint8)

dic = SubsetDIC({
    "dic": {
        "radius": 5,
        "spacing": 1,
        "cutoff_diffnorm": 1e-6,
        "cutoff_iteration": 50,
        "subsettrunc": False,
    },
    "seeds": {
        "method": "manual",
        "n_seeds": 1,
        "manual_positions": [(512, 128)],
    },
    "strain": {
        "enabled": True,
        "radius": 15,
    },
    "border": {
        "bcoef": 5,
        "interp": 5,
        "extrap": 5,
    },
})

result = dic.run(ref, cur, roi)

print(result.success)
print(result.u.shape, result.v.shape)
print(result.valid.sum())

API

The main entry point is:

from subsetdic import SubsetDIC

dic = SubsetDIC(config)
result = dic.run(ref_img, def_img, roi_mask=None)

Inputs:

• ref_img: reference image, 2D array.
• def_img: deformed/current image, 2D array with the same shape.
• roi_mask: optional 2D mask. Non-zero pixels are included in DIC.

DicResult fields:

• success: whether DIC produced a result.
• u, v: displacement fields.
• corrcoef: correlation residual field.
• valid: valid point mask.
• dudx, dudy, dvdx, dvdy: displacement gradients, available when strain calculation is enabled.
• points_computed: number of propagated valid points.
• regions: ROI regions generated from the mask.

B-Spline Preprocessing

SubsetDIC follows the Ncorr-style B-spline image preprocessing path. Before the C++ IC-GN solver starts, both reference and current images are converted to quintic B-spline coefficient images. These coefficient images are then packed into a QK_B_QKT lookup table.

The current-image gray value interpolation in the C++ core is evaluated from that lookup table, not from bilinear interpolation on the raw image. The reference-image gradients used by IC-GN are extracted from the same lookup table convention, so interpolation and gradient calculation share the same B-spline basis.

The pure Python version of this project is available at lbd-hfut/Subset-DIC and can be used as a reference for the B-spline preprocessing logic.

Configuration

Default parameters live in config/default.yaml.

dic:
  radius: 20
  spacing: 1
  cutoff_diffnorm: 1.0e-6
  cutoff_iteration: 50
  subsettrunc: true
  direct_seed_grid: false

seeds:
  method: sift
  n_seeds: 1

strain:
  enabled: true
  radius: 5

postprocess:
  enabled: true
  max_iterations: 8
  min_neighbors: 3
  corrcoef_threshold: 2.0

border:
  bcoef: 20
  interp: 20
  extrap: 20

Important parameters:

• dic.radius: subset radius in pixels.
• dic.spacing: grid spacing between calculated DIC points.
• dic.cutoff_diffnorm: IC-GN convergence threshold.
• dic.cutoff_iteration: maximum IC-GN iterations.
• dic.subsettrunc: whether to truncate subsets near ROI boundaries.
• dic.direct_seed_grid: solve each grid point with seed-style local search instead of region-growing propagation. This is slower but more robust for periodic or highly nonuniform displacement fields such as the star case.
• seeds.method: sift for automatic seeds or manual for fixed positions.
• seeds.manual_positions: list of (x, y) seed coordinates when using manual seeds.
• strain.enabled: whether to compute displacement gradients.
• strain.radius: radius used for gradient fitting.
• postprocess.enabled: interpolate bad displacement points from neighboring valid points before strain calculation.
• postprocess.max_iterations: maximum neighbor-growing fill iterations.
• postprocess.min_neighbors: minimum valid 8-neighbors required to fill a bad point.
• postprocess.corrcoef_threshold: points with larger correlation residual are treated as bad during interpolation.

You can also override selected DIC parameters when calling run:

result = dic.run(ref, cur, roi, radius=15, spacing=3, compute_strain=False)

Running Tests

Run the synthetic translation test:

python tests\test_dic.py

Expected output includes:

PASSED
All tests passed!

If pytest is installed, the same test can be run with:

python -m pytest -q

Running Example Cases

Run bundled cases:

python tests\run_cases.py

This runs:

• case/ring
• case/star

Generated outputs are saved under each case's result/ directory:

• u.npy
• v.npy
• corrcoef.npy
• valid.npy
• dudx.npy, dudy.npy, dvdx.npy, dvdy.npy when strain is enabled
• overview.png

The result/ directories are ignored by Git because they are generated files.

Development Notes

• Python files under src/subsetdic/ are used directly in editable installs.
• C++ changes under src/cpp/ require rebuilding with python -m pip install -e ..
• The extension module is named subsetdic._core.
• Arrays passed to the C++ core are stored in column-major/Fortran layout to match the Ncorr-style indexing.
• reference_code/ contains the original Ncorr source used to check algorithm behavior and edge cases.

License

See LICENSE.