2D pattern matching
Finds and locates a part in the image by comparing it with a model, returning its position and orientation.
Products
VisionSuite
The machine vision library we develop in-house: 2D pattern matching, DXF comparison, measurement, calibration, camera handling and 3D vision. Not a project delivered and closed, but a set of modules that every Tielogic system can use without rewriting them from scratch.
DXF comparison
Aligns the profile detected in the image with the part's DXF drawing and measures the deviation, point by point.
Measurement
Geometric primitives, dimensions and form tolerances to read sizes from a calibrated image.
Calibration
The camera model that converts pixels into millimetres: the precondition for any measurement that has to be a real dimension, not just a pixel count.
Cameras
Acquisition from several sources — GigE/USB3 cameras, image folders, recordings — behind a single interface, so the same code runs live or on a replay.
3D
Point clouds, meshes and pose estimation for the cases where two dimensions are not enough.
The applications we build for our customers — bespoke software and our subscription products — rest on these modules instead of code written from scratch every time: a fix or an improvement made for one system is available for the next.
Measured comparison
Against a reference library, on the same task
We compare VisionSuite with one of the world's leading commercial vision libraries in two ways: on an in-line fabric inspection, with production images, and on a test bench with synthetic images, where the right answer is known. Times are ratios: 2× means VisionSuite takes twice as long.
| Area | Time relative to the reference library | Results |
|---|---|---|
| In-line fabric inspection (real case: thresholds, areas, polygons) | 356 ms vs 221 and 573 ms vs 362 per piece (≈ 1.6×) | no difference over 4,800 analyses, 352 garments, 32 images |
| Basic operators (57, images from 1 to 23 megapixels) | 48 of 50 within 3×; 28 on par or faster on average | identical on 52 operators; 5 differ by definition, not by error |
| Caliper measurement (lines, circles, positions) | 1.2× to 2.4× | 18 comparisons, no deviation |
| 2D pattern matching (6 scenes with known pose) | 2.6× to 3.2× with one part, 4.7× to 7.0× with four | all instances found; position error ≤ 0.09 px (reference ≤ 0.10 px), angle error ≤ 0.22° (reference ≤ 0.07°) |
| Comparison with the DXF drawing (6 scenes: part straight, rotated, with a 0.30 mm dent) | 2.2× to 6.2× | part found in every case; dent measured within 0.013 mm |
Real case measured on 28 September 2026, before the latest parity fixes; test bench repeated on 5 and 7 October 2026 on a single thread. For operators the worst-case ratio counts; those under one millisecond are not ranked.
On the real case VisionSuite takes about 1.6 times as long as the reference library and, after the subsequent fixes, shows no difference in results. The first version, before a series of fixes, was 7 to 10 times slower.
The largest gaps remain in pattern matching with several parts in the scene and in DXF comparison with a rotated part: that is the work in progress. Among the basic operators two still exceed three times, a contrast-enhancement filter and line-region generation, and there too the result is the same.
Want to see it on your part?
VisionSuite is the engine behind our vision systems: if you have a similar problem, we start from there rather than from scratch.