Image physics comes before the neural network

Computer-vision projects often begin with model selection: teams compare architectures, size GPUs, and order annotation. Yet a camera does not see a “defect” as a business concept. It records light reflected from a surface. If a scratch, dent, or incorrect texture cannot be separated from a normal product in the pixels, a more complex neural network will merely process a poor signal at greater cost.

The case of Czech company 24 Vision suggests a more pragmatic order. The company is a small enterprise with 10–49 employees that develops machine-vision systems for manufacturing inspection. To move from image-processing rules to AI-based defect classification, it needed more than algorithms. Two basic questions had not been answered: which lighting would consistently reveal defects across different materials, and which acquisition setup would produce a representative dataset.

Instead of immediately buying a laboratory and training a model, the company used a Test Before Invest service at EDIH-DIGIMAT. The output was not a finished turnkey AI system. It was a validated image-acquisition chain, a dataset, and a documented method. For a small business, this intermediate result can be more valuable than an early demo that reports high accuracy on convenient photographs.

What the laboratory actually tested

The experimental workstation allowed the team to change sample holders, camera position, a polarising filter, and the geometry and type of illumination. Linear, planar, spherical, and point sources were used in radial and tangential configurations. Samples were captured in ultraviolet, visible, and near-infrared spectral ranges.

According to the published case, only the visible range reliably highlighted defects across all samples under study. This is not a universal law for every factory. It is a validated conclusion for a particular set of materials, finishes, and colour variants. Its business value is that the team narrowed the design space before purchasing a production system and eliminated unnecessary investigation of UV and NIR imaging chains.

The engagement also produced a representative image dataset and documented an operating chain: a CCD/CMOS sensor, Mono8 or Mono12 conversion, and a target theoretical detection error below one pixel. That last figure describes a technical baseline for the resulting method; it is not evidence of an error-free sorting rate on a live production line.

On its own website, the vendor separately claims reliability above 99% for its commercial system. That is the company’s product claim and cannot automatically be transferred to a new material, line, or defect class. Each customer project needs its own acceptance criteria and trials on the customer’s flow.

Why ordinary photographs do not automatically become a dataset

An industrial-vision dataset must describe not only the parts but also the acceptable operating conditions. It should cover:

  • materials, colours, coatings, and geometric product variants;
  • normal process variation without a defect;
  • real defects of different sizes, locations, and severity;
  • shifts, raw-material batches, dirty optics, and ageing light sources;
  • part position, vibration, conveyor speed, and exposure;
  • the inspector’s reference decision and reason for rejection;
  • the version of the camera, lens, lighting, and exposure settings.

The widely used MVTec AD research dataset contains more than 5,000 images across fifteen categories, with normal samples and more than seventy anomaly types supported by pixel-precise annotations. It is valuable for comparing methods. However, its licence restricts commercial use, and a laboratory benchmark cannot replace data from a specific production line.

An independent CVPR 2025 analysis of industrial anomaly detection also highlights the gap between academic tests and production. Controlled datasets with artificial defects do not capture the full diversity of real operating conditions, and strong benchmark results can degrade after deployment. For managers, the conclusion is straightforward: a proprietary acceptance set matters more than an attractive model-accuracy table.

A minimum pilot architecture

It is useful to divide the pilot into physical and software loops.

The physical loop:

  • fixes the part or passes it through a repeatable inspection zone;
  • uses an enclosure and selected lighting geometry to suppress ambient light;
  • locks the camera, lens, distance, and exposure;
  • connects the frame to a product, batch, and timestamp through a trigger;
  • uses a reference sample to detect focus, illumination, or exposure drift.

The software loop:

  • stores the unchanged source frame together with metadata;
  • lets the model return a class, defect region, and confidence;
  • applies business thresholds for the specific product;
  • sends borderline cases to a person rather than directly to scrap;
  • records the inspector’s decision as feedback;
  • monitors review rate, false rejects, and missed defects.

A local deployment is particularly appropriate when images reveal product geometry, process parameters, or supplier defects, and when latency is tied to the line cycle. The camera and inference machine can operate on the shop-floor network without sending images to an external API. Locality alone does not solve access control, however: source frames, inspector decisions, and defect reports remain sensitive manufacturing data.

How to accept a pilot without fooling yourself

A single “99% accuracy” figure is insufficient. If normal products greatly outnumber defects, a system can report high overall accuracy while still missing a rare critical fault. Before the experiment begins, agree on at least:

  • recall for each critical defect class;
  • false-reject rate by product, batch, and shift;
  • the share of frames routed to manual review;
  • p95 time from trigger to decision;
  • robustness on new batches, colours, and after line maintenance;
  • cost per accepted decision rather than per processed frame;
  • a procedure for stopping automatic rejection when drift is detected.

The test set should be separated by batch or time period rather than by randomly mixing adjacent frames. Otherwise, nearly identical images from one production series may appear in both training and validation, producing an inflated estimate. The final run must use data that the team did not use to tune the light, threshold, or model.

Economics: buy the result of the experiment, not the entire stack

The 24 Vision case does not disclose the cost of the work or realised return on investment. It is therefore appropriate to discuss only the structure of the decision. At an early stage, a company needs a budget for samples, a test rig, engineering time, annotation, and validation—not necessarily its own multispectral laboratory.

A model calculation for the pilot can use these terms:

  • missed-defect cost = number of misses × average loss;
  • false-reject cost = incorrectly rejected units × marginal unit cost;
  • manual-inspection cost = inspector minutes × fully loaded rate;
  • solution cost = pilot, equipment, integration, operation, and revalidation;
  • benefit = reduction in the first three categories minus solution cost.

The model must include changeover for a new product, lighting replacement, protective-glass cleaning, image retention, and process-engineer time for disputed cases. If those items are omitted, a demonstration rig will look more attractive than a production system.

A practical next step

Choose one station, one product type, and two or three economically material defects. Over two weeks, build a matrix of material, surface, defect, lighting, and camera position without training a complex model. The first stage should prove that the defect is reproducibly visible and measurable.

Then lock the acquisition chain, collect separate training and test batches, and run a simple baseline model. Only after the physical signal and the acceptance set are stable should the company compare heavier architectures, GPUs, and local deployment options.

The management lesson is simple: in computer vision, data quality begins not with the annotation vendor but with lighting, optics, and a repeatable process. The neural network should be the final amplifier of a signal that is already visible—not an expensive attempt to guess what the camera never captured.