Software-Defined Lighting: One Head, Switching Between Dark-Field and Coaxial Illumination

Every time a single object passes in front of the camera on an inspection line, its surface is often not made of just one material. A part where a matte-molded section and a nickel-plated terminal section sit side by side, or a PCB assembly where solder mask and exposed copper pads are intermixed — a target whose reflectivity varies sharply from region to region is closer to the norm than the exception.

The problem is that on a target like this, inspecting under a single fixed light means that the moment a defect comes into sharp view in one region, another region becomes overexposed or loses contrast. Re-tuning the lighting angle on-site stops the line, and pushing ahead with inspection under a forced compromise on a single lighting condition lets missed detections quietly accumulate. False detections can at least be caught on re-inspection, but a missed detection ships as-is.

An approach that has recently drawn industry attention is making the lighting head itself multi-channel, and switching the angle-diffusion combination needed for each inspection recipe entirely in software. This article organizes what actual optical advantages this Software-Defined Lighting structure offers, together with the interface and algorithm trends that support it.

Why multi-channel lighting is needed: one target, several reflectance behaviors

Dark-field and coaxial illumination are opposite in principle from the start. Dark-field fires light obliquely at a low angle (typically in the 0–45° range), so that a smooth surface returns no reflected light to the camera, and only a broken surface point — a scratch or a crack — scatters light up into the lens. Coaxial epi-illumination, by contrast, sends light down through a beam splitter exactly parallel to the optical axis, so that a mirror-smooth surface returns a uniform specular reflection straight back to the camera, while a defective point scrambles that reflection.

On the same target, dark-field is favorable for the matte region while coaxial epi-illumination is favorable for the high-gloss metal region. Physically splitting the two lights into separate setups adds inspection stations and lengthens cycle time. A structure such as the OptiX imaging system recently introduced by UnitX — integrating 32 independently controlled light-source channels into a single head and switching among grazing-ring (dark-field), coaxial (bright-field), dome-diffuse, and backlight states within 100 ms — is presented as a representative example of Software-Defined Lighting[1]. The key takeaway is that storing the optimal lighting condition per defect type as a software recipe makes it possible to inspect several materials in sequence at the same station.

The interface that supports switching speed: the role of GigE Vision 3.0

Switching lighting quickly doesn’t help if images can’t be pulled in just as quickly — the cycle-time gain gets cancelled out. GigE Vision 3.0, officially released by the Association for Advancing Automation (A3), adopts the RoCEv2 (RDMA over Converged Ethernet v2) protocol, bypassing the CPU in image-packet processing to cut latency and raise bandwidth utilization[2]. In a multi-recipe inspection that cycles through several lighting states while continuously capturing frames, this kind of low-latency interface is where the bottleneck gets resolved, carrying the lighting-switch speed (at the < 100 ms level) through into actual throughput. The multi-camera GigE Vision 3.0 demo shown at Automate 2026 (Chicago) is also a sign that industry adoption in this direction is underway[2].

Edge inference standing in for recipe selection

As the number of lighting recipes grows, a process is also needed to judge "which recipe should apply to the part passing through right now." In the 2026 industrial-vision edge AI landscape, M.2-form-factor acceleration modules have settled in as a standard form factor, and configurations running inference at sub-100 ms latency and sub-10 W power, without occupying the host system’s RAM, have been reported[3]. For defect classification itself, lightweight variants of the YOLOv8/YOLOv10 family (e.g., parameter reduction via ghost modules, multi-scale feature fusion) have been proposed in numerous 2026 studies[4], and it is also worth noting an August 2026 comparison study finding that, on noisy low-light images with insufficient training data, classical filters such as Median and Gaussian actually remain competitive on a cost-to-performance basis for small datasets[5]. In other words, the practical takeaway is to handle both recipe selection and defect classification on a single edge board, while balancing the filtering stage between classical and learning-based methods depending on data availability.

Core framework matching table (example setup)

The table below is an example Software-Defined Lighting setup assuming a PCB connector assembly with a mix of matte-molded and high-gloss terminal regions. Actual parameters vary by target material and defect type, so confirmation is required before finalizing any spec.

Category	Item	Value / Spec
① Minimum detectable defect size	Solder bridge/micro-crack baseline	15 µm (example baseline, requires re-verification per target)
② Optical setup — lighting	Channel configuration	3 channels — dark-field (low angle 10–15°) + coaxial epi-illumination + dome diffuse, recipe switching < 100 ms [1]
② Optical setup — lens/WD	Lens type and working distance	Semi-telecentric lens, WD 90 mm secured (including clearance for housing/lighting-head interference)
② Optical setup — camera	Sensor resolution reference	Reference case of an industrial global-shutter sensor around 24.5 MP, pixel size on the order of 2.74 µm [6]
③ Algorithm parameters	Defect classification model	Lightweight YOLO family (with ghost modules applied), example input resolution 640×640 px [4]
③ Algorithm parameters	Pre-processing filter	Median/Gaussian preferred under roughly 50–100 training images; learning-based methods used alongside once more data is available [5]
③ Algorithm parameters	Interface	GigE Vision 3.0 + RoCEv2, supporting continuous capture across multiple recipes [2]

What this matching table shows is that the decision to add lighting channels doesn’t end at optical design — interface bandwidth and edge-inference latency need to be designed together for the cycle-time gain to actually materialize. Advancing only one of these axes leaves the others as a bottleneck.

A reservation on materials with difficult reflectivity

On materials with heavy diffuse-reflection variation, such as high-gloss metal or a wet-coated surface, contrast can shift significantly with just a few degrees of change in lighting angle. For these materials, rather than applying the parameters in the matching table above as-is, a procedure of first checking each lighting channel’s response on actual samples is required. Before sample testing, it cannot be confirmed that a specific lighting combination is always effective on a specific defect.

Field Checkpoints

WD (working distance) secured: a multi-channel lighting head often has a larger housing than a single-light setup, so advance confirmation is needed that the lens-to-target working distance does not physically interfere with the lighting head.
Confirm by actual measurement, including interface bandwidth, that recipe-switching time does not become a bottleneck against the actual line cycle time (units processed per second).
If a noise filter is applied on a learning basis in low-light, low-contrast regions, confirm that the available amount of training data meets that method’s minimum requirement.
For heavily diffuse-reflecting regions, do not finalize the lighting condition before sample testing.

References (Sources)

[1] UnitX, “Dark-Field Illumination for Machine Vision in 2026” — unitxlabs.com
[2] Basler AG / Automate.org, “GigE Vision 3.0 & RoCEv2 Protocol” — baslerweb.com, automate.org
[3] ProAICraft, “Computer Vision Edge AI News in 2026” — proaicraft.com
[4] Nature Scientific Reports, “Steel surface defect detection algorithm based on improved YOLOv10 (LAM-YOLOv10n)” — nature.com
[5] AIP Advances (August 2026), “Evaluation of noise reduction methods in digital images” — pubs.aip.org
[6] Sony Semiconductor Solutions, IMX925 industrial global-shutter CMOS sensor — sony-semicon.com