Inline Weld Seam Inspection
Vision-based defect grading in 38 ms; coverage from 2.5% sampling to 100% at line rate.
Read case study
Capability
Computer Vision Development in Australia
Optics, illumination and models treated as one system - because most vision projects fail on lighting, not on the network. We design the camera choice, the light, the framing and the model together, then measure them together on the line.
Scope
What computer vision covers - and what it doesn't.
Computer vision is the discipline of turning pixels into decisions. We treat it as a co-design problem: sensor, optics, illumination and model are one design decision, not four - because the wrong lens or the wrong angle makes a great model unusable.
In scope
- Detection, classification, segmentation and keypoint models
- Camera and lens selection, illumination design (LED bar, dome, structured)
- Framing, calibration, and mechanical mount design
- Line-scan and area-scan industrial cameras with GigE / USB3 / MIPI
- Model training with production-representative data
- Deployment on Jetson, Coral, industrial PCs, and MCU with vision AI
- Validation against known-good and known-defect samples
Honestly out of scope
- Pure algorithm research with no deployment target
- Off-the-shelf machine-vision reseller work (we design, we do not resell)
- Full-time on-site inspection line operations
- Turnkey mechanical machine-vision cell builds (we integrate)
Outcomes
What you get out of it.
A vision system that works is one that runs unattended for months at accuracy that holds. What our vision programs consistently deliver:
Accuracy that holds against production variance
Models trained on data that spans real illumination changes, part variation and edge cases. Not the curated benchmark set - the messy production one.
Illumination designed for the class of defect
Dark-field for surface scratches, bright-field for dimensional, structured light for 3D. The light choice is made before the model is trained, not after it disappoints.
Inspection at line rate, not sampled
Vision cells sized to grade every unit, not one in forty. Our weld-inspection system moved coverage from 2.5% sampling to 100% at 45 ms per weld.
A path to update as products change
Retraining pipelines designed for the day a new SKU arrives. Data collection tools included in the handover.
Process
How a computer vision program runs here.
Four phases with a real deliverable at each gate - you always know what you paid for and what ships next.
PHASE 01
Discover (paid week)
We start from the constraint that binds - power, latency, thermal, certification - and design backwards from it. You leave with a written architecture, a budget range and the risks named, whether or not we build it.
PHASE 02
Design
Schematics, mechanical and firmware architecture proceed in parallel. High-risk blocks get simulated or breadboarded before the full layout commits.
PHASE 03
Build
Iterative revisions against real bench and field testing. You see every revision, not just the last one. Integration is continuous, not a phase.
PHASE 04
Deploy
Pilot in the field, closure with the contract manufacturer, production test procedures, and a commissioning-grade handover pack.
Technologies
What we build on - chosen per constraint, not per preference.
The platforms we reach for most. If a project needs something not on this list, we say so - the tool is chosen for the constraint, never to fit our habits.
YOLOv8/v11Detectron2MMDetectionSegment AnythingOpenCVHALCONCognex VisionPro (integration)GigE Vision (Basler, FLIR, IDS)USB3 VisionMIPI CSI camerasNVIDIA DeepStreamExecuTorchTensorRT
Industries
Where computer vision we ship runs today.
Deliverables & IP
What ships to you at handover.
Every vision program hands over: trained models with model cards, camera and lens specifications with justification, illumination design and photometric measurements, mechanical mount drawings, calibration procedures with sample images, inference wrapper code, validation dataset (or references to it), and a written report of accuracy on the deployment-representative test set. All foreground IP transfers on payment.
Case studies
Programs we shipped in this space.
Every entry links to the full case study - constraints, what we built, and what it measured afterwards.
Substation Perimeter Classifier
Weather-hardened vision system distinguishing wildlife from intruders on 24/7 duty.
Read case study
Grain Protein Sorting Classifier
High-throughput classification for agricultural grading with tight ROI budget.
Read case study
Compliance
Privacy-aware and safety-honest.
Vision systems that see people trigger the Australian Privacy Principles by default - we design with data-minimisation as the starting point (process on-device where possible, retain only what is required for a documented purpose, obtain consent where the deployment context requires it). Where the deployment is workplace surveillance, we work with the client to align to Fair Work and state surveillance-notification requirements.
Safety-critical vision (autonomy, medical, security) gets the fallback path designed explicitly - what happens when confidence drops, when the scene occludes, when illumination changes. We document known failure modes as first-class deliverables, not as an appendix.
FAQ
Computer Vision Development in Australia - straight answers.
Vision inspection cell on existing line: AUD $25,000-$80,000. Full custom vision product with model training and deployment: AUD $50,000-$200,000+. Recurring model retraining engagements from AUD $6,000/quarter.
For a defect-detection model, we typically need 200-2,000 labelled examples per class, weighted toward the defect classes and covering the range of real variation. We design the data-collection plan in week one, before model architecture is finalised.
Yes, if they are appropriate - and we say if they are not. A camera that is wrong for the class of defect (wrong resolution, wrong exposure control, wrong lens) will cap accuracy no matter how good the model is; that conversation happens in the discovery week.
Yes - lighting is often the single biggest determinant of vision-system success and we treat it as a core design choice, not an afterthought. Dark-field, bright-field, coaxial, structured, backlight - each solves a different class of problem.
Yes - most of our deployments run on Jetson-class embedded compute, industrial PCs at the line, or increasingly on MCU-class hardware with vision accelerators. See <a href="/services/edge-ai-development/">edge AI</a> for target selection.
You do. Model weights, training dataset (subject to any client-data agreements), calibration procedures, and integration code all transfer on payment.
Why Incendio
One team, no seam between vendors.
Vision projects that pass on the workstation and fail on the line usually failed at co-design - the lens or the light or the mount was wrong for the actual scene, and no model recovers from that. Because we hold the mechanical, illumination and model design together, our vision cells reach their accuracy targets on the line, not in the demo.
Related practices: edge AI, robotics, industrial automation, embedded systems.
Start
Tell us the constraint that worries you most.
A latency budget, a power budget, a certification date. We reply within one business day - and we’ll say so if we’re not the right team.