Application example
Six-camera assembly verification
- Challenge
- Images arrived late on a shared network switch.
- Setup
- Embedded controller with six PoE ports and jumbo frames.
- Result
- Deterministic image arrival and results within cycle time.
Components · Computing
Vision Controllers & Computing
The controller runs the inspection: it acquires images, executes the vision program, makes decisions within the cycle time and talks to the line. Sizing it correctly prevents dropped images, late results and expensive upgrades.
Bộ điều khiển vision chạy chương trình kiểm tra: thu ảnh, xử lý, ra quyết định trong thời gian chu kỳ và giao tiếp với dây chuyền.
How it works
From camera to decision within the cycle time
Every inspection follows the same chain: trigger → exposure → image transfer → processing → result → output. The controller must complete this chain for every part, every time, within the line's cycle time — including peaks when several cameras fire together.
Processing load depends on image size, number of images, the tools used and whether deep learning runs on the system. Rule-based tools usually run well on modern multi-core CPUs; deep learning inference often benefits from a GPU or AI accelerator.
Reliability matters as much as speed: industrial hardware, controlled operating-system updates, image storage planning and a clean electrical installation keep the system running for years.
Types
Computing platforms in detail
Open each type for how it works, where it fits, its limitations and the specifications to check.
01 Embedded vision controllerCompact, fanless controller with built-in camera ports and I/O.
How it works. Industrial computer designed for vision: PoE camera ports, isolated I/O, light-control outputs and pre-installed vision runtime.
Best for
- Stations with 1–8 cameras
- Machine builders needing a compact cabinet unit
Limitations
- Fixed expansion options
- Limited GPU power on smaller models
Key specs to check
- CPU class
- Number of PoE / USB3 ports
- I/O count
- Mounting (DIN-rail, wall)
02 Industrial PC + frame grabberMaximum flexibility and performance.
How it works. Box or rack industrial PC with expansion slots for frame grabbers (CoaXPress), network cards, GPUs and I/O cards.
Best for
- High-speed and line-scan systems
- Many cameras or very large images
- Custom software
Limitations
- Larger footprint
- Requires careful thermal design in cabinets
Key specs to check
- CPU cores
- PCIe slots
- RAM
- Storage (SSD, RAID)
- Power supply
03 GPU edge computerRuns deep-learning inference on the line.
How it works. Industrial computer with GPU or AI accelerator for fast neural-network inference and, optionally, on-site training.
Best for
- Deep-learning defect detection
- Classification of variable defects
- OCR on difficult text
Limitations
- Higher power and heat
- GPU models and drivers must be managed
Key specs to check
- GPU / accelerator class
- Memory
- Cooling
- Supported frameworks
04 Frame grabberHigh-bandwidth acquisition card.
How it works. PCIe card that receives images from CoaXPress (or legacy) cameras with deterministic timing, buffering and I/O.
Best for
- CoaXPress cameras
- Line-scan systems
- Very high data rates
Limitations
- Requires a PC with suitable PCIe slots
Key specs to check
- Links and speed
- PCIe generation and lanes
- On-board memory
- I/O and encoder inputs
05 PoE network cards & switchesReliable GigE camera networks.
How it works. Multi-port PoE network cards give each camera its own link; managed industrial switches support larger installations.
Best for
- Multi-camera GigE systems
- Long cable runs
Limitations
- Bandwidth per port must be respected
Key specs to check
- Ports and speed
- PoE power budget
- Jumbo frame support
06 I/O & trigger modulesTiming between motion, light and camera.
How it works. Encoder inputs, trigger distribution, light strobe control and digital outputs for reject mechanisms.
Best for
- Line-scan and encoder-based systems
- Synchronizing several cameras and lights
Limitations
- Wiring and timing must be documented
Key specs to check
- Input/output count
- Response time
- Isolation
Compare
Sizing guide (starting point)
| System | Typical platform | Notes |
|---|---|---|
| 1–2 cameras, rule-based tools | Smart camera or small embedded controller | Lowest cost and footprint |
| 3–8 GigE cameras, rule-based | Embedded controller with PoE ports | One port per camera |
| Deep-learning inspection | GPU edge computer | Size GPU for image size and cycle time |
| Line-scan / CoaXPress | Industrial PC + frame grabber | Plan PCIe lanes and storage speed |
| Many stations, central data | Controllers per station + server | Store images and results centrally |
Final sizing should be confirmed with a test of the real vision program and image sizes.
Specifications
Key specifications explained
| Specification | What it means | How to choose |
|---|---|---|
| CPU | Number of cores and generation. | Parallel cameras and tools benefit from more cores. |
| GPU / AI accelerator | Hardware for neural-network inference. | Required for most deep-learning workloads at line speed. |
| Camera ports | PoE GigE, USB3, frame grabber links. | One dedicated port per camera for GigE. |
| RAM | Working memory for images and buffers. | Large images and many buffers need more memory. |
| Storage | SSD capacity and write speed. | Plan image retention: images per day × size × days. |
| I/O | Digital inputs/outputs, encoder, light control. | Match triggers, rejects and handshakes. |
| Operating environment | Temperature range, fanless, vibration. | Match the cabinet conditions. |
| Power | Supply voltage, UPS support. | Protect against power dips and unsafe shutdowns. |
Engineering tools
Calculate before you choose
Quick estimates based on standard formulas. Always confirm with a test on real parts.
Data Rate Calculator
Estimate the data each camera sends and the total the controller must receive.
Selection guide
Which platform for which task?
| Requirement | Recommended | Why / note |
|---|---|---|
| Single inspection, simple tools | Smart camera | No separate computer. |
| Several GigE cameras | Embedded controller with PoE ports | Compact cabinet installation. |
| Variable defects, deep learning | GPU edge computer | Fast inference on the line. |
| CoaXPress or line-scan | IPC with frame grabber | Deterministic high-bandwidth acquisition. |
| Factory-wide data and images | Station controllers + central server | Connect to MES or a data hub. |
Selection steps
- Count cameras, image sizes, frame rates and images per part.
- Calculate data rate and check ports and interfaces.
- Estimate processing time per image with the real tools.
- Decide on CPU-only or GPU for deep learning.
- Plan I/O, triggers, encoder and reject outputs.
- Plan storage, backup and network integration.
- Specify environment, mounting and power protection.
Examples
Typical application examples
Illustrative setups showing how the components are combined in practice.
Application example
Deep-learning surface inspection
- Challenge
- CPU inference was too slow for line speed.
- Setup
- GPU edge computer with optimized inference.
- Result
- Inference time reduced to fit the cycle with margin.
Application example
Line-scan web inspection
- Challenge
- Very high continuous data rate.
- Setup
- Industrial PC with CoaXPress frame grabber and fast NVMe storage.
- Result
- No lost lines and full defect image archive.
Common mistakes to avoid
- Sharing one network port among several GigE cameras.
- Not enabling jumbo frames and correct network settings.
- Underestimating image storage needs.
- Uncontrolled operating-system updates on production systems.
- Ignoring cabinet temperature for fanless or GPU systems.
- No UPS — unsafe shutdowns can corrupt storage.
FAQ
Frequently asked questions
Do I need a GPU?
For rule-based tools, usually not. For deep learning at production speed, a GPU or AI accelerator is usually recommended.
How much storage do I need?
Multiply images per day by image size and the number of days you must keep them; compressing OK images and keeping all NG images is a common strategy.
Can one controller run several stations?
Yes, if ports, processing time and I/O allow it — but separate controllers make stations independent and easier to maintain.
Windows or Linux?
Both are used. The choice depends on the vision software, IT policies and maintenance skills; what matters most is controlled updates and stable drivers.
Continue exploring
Related components
Vision SoftwareNo-code tools, SDKs, deep learning, 3D processingIndustrial CamerasArea-scan, line-scan, interfaces & special sensors3D VisionLaser profilers, structured light, stereo, ToFIndustry solutionsInspection points by industry
Not sure which configuration fits?
Send us your parts and requirements — our engineers test real samples and recommend a setup.