VisionControl
This sheet summarizes value, pricing and use cases. From here you can move to the dedicated areas or to the client portal for the next action.
Visual inspection does not stop at recognizing the defect: VisionControl takes the inspection to a line that is truly running, with acquisition on trigger, result within the cycle time and automatic rejection. Runs inside the factory perimeter: production images do not leave the premises.
What it solves
- Sampling inspection lets defects pass from one sample to another, and 100% human-based inspection cannot keep up with the line pace
- Between the successful lab test and the self-scraping installation lies all the real work: synchronization on the trigger, cycle time, uncontrolled lighting, scrap command
- Defects change when the process changes: a once-trained model degrades and no one notices until returns arrive
- The suites from major vendors are paid for twice, development license plus runtime license, and no one publishes the prices: you can't get an idea of the cost without opening negotiations
VisionControl is an artificial vision system for in-line quality control: it acquires the piece image on the machine trigger, decides within the cycle time and commands the rejection of non-conforming pieces. Runs on-premise, inside the factory perimeter: production images do not leave the premises.
The differential is not about recognizing the defect — any vision library can do that — but going into production on a line that's running: synchronized acquisition, decision within the cycle time, automatic rejection, and a model trained on that customer's actual defects, not on a lab dataset.
What it does
| Area | What it does |
|---|---|
| Field Acquisition | Synchronized shot on the GPIO trigger of the machine, management of single and group cameras (grid, multi-view, composite) |
| Inspection Engine | Comparison with the master part, geometric thresholds, pass/fail outcome within cycle time |
| Anomaly Detection | Trained models on real customer defects (PatchCore, PaDiM, EfficientAd) for defects that cannot be described by a rule |
| Inspection Plugin | 21 runtime loaded plugins: adds a control without touching the engine |
| Reject Pipeline | Command to reject towards the machine, with real or simulated GPIO backend for tuning |
| Light Robustness | Gray-world white balance, CLAHE and regions of interest, to handle uncontrolled ambient lighting |
| Temporal Ghosting | Master/acquisition overlay and defect heat map, to understand where the line is failing |
| Trend and Drift Analysis | Detection of drift on outcomes with automatic email alert, threshold per camera and cooldown period |
| Active Learning | The operator marks false positives, false negatives and borderline cases: the model corrects itself based on real-line defects |
| Process Data Import | Excel, CSV and external databases, to correlate inspection outcome with production data |
| Traceability | Append-only log of operations and history of outcomes |
- Feasibility study on customer samples — performance is measured on real pieces before committing, not on catalog images.
- Training the model on actual defects — this is the work that opens projects and what separates a successful proof of concept from an operational installation.
- Line calibration — synchronization on trigger, cycle time, illumination, reject command.
- Monitoring and re-calibration — defects change as the process changes: the model needs to be re-verified over time.
Architecture
- Backend: Python 3.11, FastAPI, SQLAlchemy 2 async, Alembic, Pydantic v2. - Frontend: React 18, Vite, TypeScript, Tailwind, shadcn/ui. - Asynchronous processing: Celery workers for inference and import. - Plugins: Python modules loaded at runtime from the plugins folder. - Data: PostgreSQL.
Requirements
- On-premise distribution at the factory; PostgreSQL 14 or higher.
- Industrial cameras with trigger available (GPIO), or network video source for pure monitoring configurations.
- For reject command: Digital I/O to the machine. In absence, the simulated GPIO backend allows tuning everything else.
- Access to sample parts, conforming and defective, for model training: it is the prerequisite that determines system performance.
Reports & dossiers
- Report di scartoPDF
Documenta il pezzo scartato con l'immagine e il riquadro del difetto sovrapposto: la prova di cosa e' stato scartato e perche'.
- Cruscotto di active learningVista applicativa
Falsi positivi, falsi negativi e casi borderline segnalati dall'operatore, per correggere il modello sui difetti della linea reale.
- Alert di derivaEmail
Segnalazione automatica quando l'andamento degli esiti si discosta dalla norma, con soglia per telecamera e cooldown.
- Storico degli esiti e registro operazioniVista applicativa
Esiti di ispezione nel tempo e registro append-only delle operazioni, per la tracciabilita' interna.
Frequently asked questions
Su quali ambiti lavora VisionControl?
Two with very different requirements. Biomedical: posture analysis and skin screening with the characteristics ABCDE. Industrial: detection of surface and structural defects on production lines.
Come decide se un pezzo è difettoso?
Con due meccanismi che si completano: il confronto a livello di pixel con un pezzo di riferimento, e il rilevamento di anomalie con modelli addestrati sul buono anziché sul difettoso. L'acquisizione è sincronizzata sul segnale della macchina e lo scarto può essere comandato in automatico.
Funziona se la luce in reparto cambia durante la giornata?
It is one of the issues that render a vision system unusable, and VisionControl tackles it with white balancing, adaptive contrast equalization, and delineation of the area to be examined. It does not work miracles on incorrect lighting, but tolerates the normal variation in a warehouse with windows.
Can it be corrected when it makes a mistake?
Yes, and it was planned from the beginning. The operator reports errors, and the system collects them in a view of false positives, false negatives, and uncertain cases: it is from there that one understands whether the model needs to be retrained or if something has changed in production.
Can the measurements be cross-referenced with production data?
Yes. VisionControl imports tables from Excel, CSV, and external databases, so a measurement derived from an image can be placed next to process or master data. This is how a defect becomes a correlation rather than a count.
Use cases
Controllo di superfici verniciate e cromate su linea automotive
Ispezione dall'esterno di prodotti verniciati e cromati, il caso piu' ostile per la visione per via dei riflessi: bilanciamento del bianco gray-world, CLAHE e aree di interesse per reggere la luce ambientale, confronto con il master e rilevamento anomalie addestrato sui difetti reali della linea.