Optical algorithms
Custom 2D/2.5D/3D imaging design
Optical algorithms, image-processing algorithms and vision deep learning are unified in one platform. Labelling, training, inference and deployment happen without code, and built-in 2D/2.5D/3D modules detect defects against complex backgrounds.
- Detection accuracy
- Up to 99.99% · application-specific
- Detection precision
- From 1 μm · depends on the optical configuration
- Inspection speed
- Up to 200 units/s · depends on workpiece and hardware
- Minimum training samples
- Typically 5; from 1 for large parts
- Dimensions
- 2D / 2.5D / 3D / X-ray / infrared
- Acceleration
- NVIDIA · Intel · Huawei · FPGA
- Integration
- MES · PLC / Linux · Windows
Handdle AI
Three algorithm layers, one platform
Inspection begins in the optics and ends in the network. Holding all three layers in one platform means performance does not leak between them.
Image processing
Geometric fitting, colour processing, contour analysis
Vision deep learning
Defect detection, classification, measurement, recognition
Handdle AI
Core functions
Eight modules combine into a single inspection pipeline.
Detection
Automatic object location — part presence, positional deviation and foreign matter in real time.
Segmentation
IOU up to 99.9%
Application-specific
Pixel-level defect segmentation yielding position, area and shape.
Classification
Automatic classification by defect type, including grading against complex textures.
One-class training
Trains on good samples alone, detecting anomalies without NG data.
Dimensional measurement
Length, width, angle and gap measured automatically at pixel precision.
3D detection
Height, depth and flatness inspection, integrated with line laser and structured light.
OCR / OCV
Character recognition and verification: print quality, typos, label conformity.
Code reading
1D/2D barcodes, QR and DPM read and linked to process traceability.
Handdle AI
Technical advantages
Lower data dependency, without giving up accuracy or speed.
Few-sample training
From 5
From 1 for large parts
Typically trains and deploys from 5 images — and from a single sample where the part is too large to bring to us.
One-class training
0 NG
Trains on OK samples only, so a new product with no defect history can be covered immediately.
Transfer learning
30+ models
Shared characteristics across same-material products are reused in high-mix environments.
Detection accuracy
Up to 99.99%
Application-specific
Multi-dimensional feature extraction with automatic filtering reduces over- and under-detection together.
Detection precision
From 1 μm
Depends on the optics · defects resolve at 2–3 pixels
Pixel-level matching of the defect region yields position and area for fine defects — resolving at 2–3 pixels where rule-based systems typically need five or more.
Inspection speed
Up to 200 units/s
Depends on workpiece and hardware
Simultaneous software and hardware acceleration sustains full accuracy at speed. Actual throughput is sized to the workpiece, resolution, model and hardware.
Broad compatibility
5 image types
2D, 2.5D, 3D, X-ray and infrared imagery; major acceleration platforms; MES and PLC integration.
Visual debugging
Heatmap
Exposes the basis for each decision, cutting tuning time and dissolving the black-box problem.
Handdle AI
Deployment flow
Training and operation run as one continuous path.
- 01
Training
Image capture → labelling → model training → model export
- 02
Deployment
Import config → configure modules → predict and infer → one-click run
Data security
Inspection data stays inside your facility
The default deployment is on-premise: training, inference and storage all run on customer infrastructure, and cloud connectivity is configured only on request.