Surface defect detection
Cracks, porosity, inclusions, scratches, missing material. Semantic segmentation or multi-class classification on a real-time stream.
CNN · UNet · YOLOv8 · TensorRT
Domain of expertise
Industrial vision and AI quality control
From camera to quality decision: embedded or remote vision systems for inline inspection, integrated with the PLC or MES. Inference < 200 ms, edge deployment (Jetson, Hailo) or industrial server, OPC-UA / Profinet connectivity.
Capabilities
What we build
Non-contact dimensional control
Measurement of critical dimensions by sub-pixel vision. Comparison to the CAD nominal or to masters, configurable out-of-tolerance alert.
Stereo calibration · Profilometry · OpenCV
Assembly inspection
Presence / absence of components, orientation, positioning, marking legibility (DataMatrix, QR, laser engraving).
Template matching · OCR · Object detection
3D vision
Laser profilometry, point clouds, measurement of volumes and complex geometries. For defects that cannot be detected in 2D.
Structured light · LiDAR · 3D reconstruction
Automatic sorting and ejection
Digital signal to the PLC for immediate ejection. Timestamped part-by-part logs for audit traceability (IATF, ISO 9001, GMP).
OPC-UA · Profinet · GPIO · REST → MES
Unsupervised anomaly detection
For cases without labelled defect data: learning the normal distribution and detecting deviations. Useful at start-up or for rare defects.
PatchCore · FastFlow · Autoencoders
Architecture
Reference stack
Acquisition → processing → inference → decision. Each building block is sized to the cadence, resolution and integration constraints of the site.
Acquisition
Camera + lighting
GigE Vision, USB3 Vision, SWIR depending on the defect type. Coaxial, grazing, backlit or multispectral LED.
Runtime
Embedded or remote inference
NVIDIA Jetson Orin / AGX, Hailo-8, industrial server. ONNX / TensorRT for latency optimization.
Decision
Business logic + thresholds
Confidence thresholds configurable per defect class. Calibrated with the quality manager to control FP / FN.
Integration
PLC & IS connection
OPC-UA, Profinet, Modbus, REST. Logs to MES / ERP. Real-time operator interface with defect heatmap.
Sectors
Industrial contexts
Designed to run in real conditions: vibration, variable ambient lighting, oil, dust, cycles < 1 s.
Plastics: injection, blow molding
Forging and foundry
Machined and ground parts
Food processing: foreign bodies
Pharma / cosmetics: packaging
Electronics: PCB, SMD components
Industrial glass and ceramics
Technical textiles and nonwovens
Results
Observed orders of magnitude
Measured on comparable deployments. They depend on the type of defect, the cadence and the quality of the initial data: presented here as benchmarks, not guarantees.
> 95 %
detection rate on targeted defects after validation
20 – 40 %
scrap rate reduction observed on instrumented lines
< 2 %
false positives after threshold calibration with the quality manager
30 – 150 ms
inference latency depending on model complexity and resolution
Approach
How we proceed
01
Vision audit
Analysis of parts, defects, cadence, mechanical constraints and lighting. Choice of the acquisition architecture. 1 week.
02
Collection & labelling
Building the dataset: 300+ images per defect class. Labelling with the quality manager. 2 to 4 weeks.
03
Prototype
Model training, threshold calibration, measurement of detection / false positive rates. Validation on a test bench. 3 to 5 weeks.
04
Production integration
PLC wiring, triggers, operator interface, load tests. Go-live on the target line. 2 to 4 weeks.
05
MLOps
Drift monitoring, real-time quality dashboard, scheduled retraining on new production data.
Get started
A scrap rate to reduce or a line to instrument?
Tell us about the context: part type, cadence, nature of the defects. We reply with an initial feasibility analysis within 48 hours.