A computer vision system is an optical instrument with software attached, and it is usually the optics that decide whether it works. Faststream engineers the whole chain — sensor and lens selection, illumination geometry, trigger and exposure, then detection, segmentation, measurement, OCR and verification — because a defect that is not visible in the image cannot be recovered by any model downstream.
Photograph real parts with real defects under several illumination geometries. Establish that the feature of interest produces contrast at all before anything else is decided.
Resolution from the smallest feature that must be resolved, working distance and field of view from the mechanical envelope, depth of field from part variation, sensor from the light available in the exposure the line speed permits.
Geometry chosen for the defect physics — low-angle for surface relief, dome for specular, backlight for silhouette and dimension, coaxial for flat specular, structured for three-dimensional form.
Encoder-locked capture, strobe timing and exposure short enough that motion blur does not consume the resolution the optics just bought.
Classical vision where it is sufficient and deterministic. Learned models where appearance varies in ways a rule cannot express.
Result within the cycle-time budget, delivered to the controller in the form it expects, with images and outcomes retained for traceability.
The single most consequential choice in an industrial vision system, and the one most often made last.
| Geometry | Makes visible | Typical use |
|---|---|---|
| Low-angle / dark field | Surface relief, scratches, edges, engraving | Scratch and dent detection on smooth surfaces |
| Dome / diffuse | Colour and print on curved or specular parts, without hotspots | Label and print inspection on cylinders and cans |
| Backlight | Silhouette, outline, through-holes, presence of opaque features | Dimensional gauging and presence checks |
| Coaxial | Flat specular surfaces without glare | Wafer, glass and polished metal inspection |
| Bar / directional | Texture and directional features | Weave, grain and seam inspection |
| Structured light | Three-dimensional form and height | Gap and flush, bead geometry, volume |
| Infrared | Thermal contrast, and imaging through some materials | In-cab monitoring, seal integrity, moisture |
Because a feature that produces no contrast in the image contains no information, and no model can recover information that was never captured. Most failed vision pilots are illumination failures — a scratch that only appears under low-angle light, a print defect washed out by a dome, a specular surface saturating the frame.
For dimensional measurement, presence and absence, code reading and most fixed-geometry tasks. Classical methods are deterministic, explainable, need no training data and do not drift. Learned models earn their place where appearance varies in ways a rule cannot express.
OCR reads characters and returns what they say. OCV verifies that the characters match an expected string and that print quality is acceptable. Packaging lines usually need OCV, because the value is in catching a wrong or illegible date code, not in transcribing it.
Usually. The constraints are mounting position, available trigger signal, cycle-time headroom and lighting control. A site survey establishes those before hardware is specified.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.