AI in machine vision, where rules are no longer enough

Name: Opales
Address: 140 rue Désiré Monnier, Lons-le-Saunier, 39000, FR
Telephone: +33 3 84 24 88 66

Surface defects that never look the same, textured materials, poorly printed markings: AI recognises what no rule describes well. Since the early 2000s, Opales has built it into VisiOpal; today, AI also helps write and test the programs.

Timeline

Two generations of AI in our inspections

The first generation inspects the parts. The second works with you on the program itself.

Generation 1

AI that inspects

2000
The first neural networks
From the early 2000s, where conventional tools reached their limits, we used neural networks in our location (pattern) and character reading (OCR) tools.
2014
Cognex ViDi in VisiOpal
Cognex's ViDi deep learning tool is built into VisiOpal: it is used as a program line, alongside the conventional tools.
2021
VisionPro Deep Learning
Cognex VisionPro Deep Learning tools join VisiOpal, with the same parameter and result conventions.
2026
Our own AI tools, ONNX format
Supervised and unsupervised AI tools developed by Opales, in the open ONNX format, run locally on the vision PC.

Generation 2

AI that programs with you

2026
AI built into VisiOpal
An MCP server built into VisiOpal lets the AI insert program lines, test them and analyse the images directly.
2026
The VisiOpal IA platform
It connects VisiOpal to AI models: our specialised VisiOpal model served by LM Studio, open models run locally or on a small dedicated server, or the connected AIs Claude and Claude Code, OpenAI Codex and Gemini CLI.

The VisiOpal IA platform

Generation 2 · VisiOpal IA

Show the area, ask the question: the AI acts in VisiOpal

The VisiOpal IA platform is connected to VisiOpal through a prompt window with selection on the image. Through the VisiOpal MCP server, the AI reads the images, inserts program lines, tests them and analyses the results.

You show and ask

In VisiOpal, the prompt window shows the image: you draw around an area or point to a detail, then ask your question in your own words.

VisiOpal IA passes it on

The platform sends the request, the area and the image to the chosen model, and keeps the whole dialogue in its log.

The model reasons

Specialised VisiOpal model, open model run locally or connected AI: it decides which actions to take.

It acts through the MCP server

It reads the image, inserts program lines, tests them, analyses the results, then answers you or asks you a question.

Opales AI request window in VisiOpal: area drawn around a DataMatrix and the AI's answer — The prompt window in VisiOpal: an area drawn on the image, the question, and the AI's answer, which reads the DataMatrix and the text "GOOD CODE".

VisiOpal IA platform window: list of requests and log of the tools called — The VisiOpal IA platform: requests received, tools called in VisiOpal, model used (here Qwen 3.5 running locally in LM Studio), tokens and duration.

Our VisiOpal model

A model prepared by Opales for VisiOpal, served by LM Studio on the vision PC or on a server in the plant.

Open models, run locally

Open language models (Qwen, Llama, Mistral…) on the PC or on a small dedicated AI server: nothing leaves the plant.

Connected AIs

The platform also works directly with Claude and Claude Code, OpenAI Codex and Gemini CLI, when the project allows it.

The MCP server is installed on the vision PC and refuses any change to the program during production.

When AI, when rules?

Each technique where it performs best

AI is not the answer to everything. A dimension is measured, a presence is checked: explainable rules, set by a tolerance, remain safer. What sets us apart: combining both in the same program.

L'IA

Variable surface defects: inclusions, stains, missing material
Natural or textured materials, where "normal" varies
Degraded markings: smudged inkjet, dot peen, faint text
Rare defects: an unsupervised model learns only what is good

Rules

Dimensional measurement, down to the micron
Presence, absence, position, counting
Barcode and DataMatrix reading
Robot guidance and part location

Both together

In VisiOpal, a conventional tool locates the part, the AI inspects the area, a rule measures the size of the defect found and decides according to your tolerance. Each step is a line of the same program, with its results visible.

The VisiOpal software

The approach

Six steps, from the first image to monitoring in production

Collection

Images taken in production, with the existing system or a test station: the real variations in light, material and position.

Annotation

Defects are marked on the images with your quality experts, who know what is good and what is not.

Training

The model is trained at Opales, then exported in the open ONNX format and run by VisiOpal.

Parallel validation

The model runs alongside the current inspection without making decisions: its verdicts are compared with yours on real production.

Going into production

The model makes the decision, combined with the conventional tools in the same program, with image archiving.

Drift monitoring

Doubtful images are kept; when production changes (material, supplier, lighting), the model is retrained.

What we need

What we ask of you, and when

Quantities depend on the defect and how much it varies: the feasibility tests make them precise. The orders of magnitude opposite help you prepare the collection.

From the feasibility study: a few good and defective parts, or images, to check that the defect is visible.
Supervised AI: images of good parts and of each defect type, generally a few dozen to a few hundred examples per defect.
Anomaly detection: images of good parts only, a few hundred; useful when defects are rare.
AI OCR: images of the marking in all its real states, sharp, faint, smudged, offset.
During validation: a person from quality to confirm doubtful verdicts.

Sovereignty

Your AI runs on your premises

Local execution

Inspection models run on the vision PC, on the processor or graphics card: no image or production data leaves the plant.

Open format

The models use the standard ONNX format, readable by other tools: nothing is locked into proprietary software.

No subscription

No cloud connection or subscription is needed to run the inspection: the line depends on no external service.

In production

AI inspections already in service

Detection and false-reject rates depend on each production: we measure them during parallel validation and give them to you before going into production.

Plastics

Capsule integrity inspection with artificial intelligence

The needDetect inclusions and missing material too variable to be described by rules, and know which mould cavity each capsule comes from.

The solution

Camera with an artificial intelligence system
Deep learning algorithm
Detection of inclusions and missing material
Traceability through the mould cavity number

AISurfaceTraceability

All industries

Reading markings with AI OCR

The needRead dates, batches and references printed by inkjet or dot peen, often faint or smudged.