AI
Raspberry Pi 5 AI: quality inspection without the cloud
A reversed connector, a missing accessory, a defect to catch: evaluate Raspberry Pi 5 and Sixfab AI HAT+ for visual quality inspection with local AI.
04 / AI
Computer vision
Measurable intelligence, in the real world
Image analysis for classification, detection and inspection: models are evaluated alongside optics, lighting and acquisition conditions.
Platforms and tools follow your context. Some examples:
Python / NumPy / C++ / TinyML / Computer vision / Edge AI
AI
Computer vision
Image acquisition and analysis for the real use case
Image acquisition
Cameras · optics · lightingSource quality
Data preparation
Labels · selection · checksRepresentative cases
Analysis & models
Classification · detectionSegmentation and inspection
Metrics & errors
False positives · missed casesReal-problem evaluation
Integration
PCs · edge · existing systemsPlatform constraints
Field verification
Lighting · variability · performanceBehaviour on real cases
The technologies shown are examples.
Engineering support within your team, with shared tools and working methods.
TECHNICAL SCOPE
Computer vision and industrial inspection
Image quality is part of the system. Observable defects, framing and variations in parts, background and lighting are defined. Separate test samples distinguish generalisation from memorisation. The interface must show results and allow uncertain cases to be inspected with sufficient context.
Activities and verification
- Prepare images and annotation criteria consistent with the inspection task.
- Select classification, detection or segmentation according to the problem.
- Measure false rejects, missed defects, latency and behaviour on unfamiliar data.
Initial decisions and integration
I clarify objects, observable defects, acquisition and expected variation with client specialists. I can work on datasets, acquisition software or inference within the existing environment. Models and thresholds are evaluated alongside the camera, optics and lighting.
A verification example
A test set can include variations in lighting, position, batches and defect-free cases. False rejects and missed defects are examined under documented conditions. Synthetic data and simulation support experiments, but transfer to real captures needs dedicated verification.
Deliverables and acceptance criteria
Pipeline, model, verification scenarios and error analysis. Synthetic data, as explored in SLX SynthForge demonstrators, can support experiments but do not replace validation on real captures.
Where to start
Bring real images, defect examples, camera and optics specifications and available time per acquisition.
Are synthetic images enough to evaluate inspection?
They can explore controlled variation. Estimating behaviour in the operating environment also needs data relevant to that environment and explicit comparison criteria.
Insights
All articles ↗↗
AI
SLX SynthForge on GitHub: AI for industrial vision
SLX SynthForge brings industrial vision to GitHub: 3D boards, synthetic data and softmax classifiers, with a C++ engine, Qt interface and Python/NumPy training.
AI
Industrial machine vision: four Ethernet cameras with Microchip and NVIDIA
Microchip’s PolarFire Ethernet Sensor Bridge Rev 2.0 supports up to four MIPI CSI-2 cameras over dual 10GbE. Learn how to assess bandwidth, latency, synchronisation and NVIDIA integration for industrial inspection.
THE NEXT STEP