AI VISION
MACHINE VISION
INSPECTION
Full-stack AI vision and machine vision electronics engineering — GigE Vision and Camera Link camera interface, multi-mode illumination design, FPGA-based image acquisition, NVIDIA Jetson and Hailo NPU AI inference, deep learning defect detection, PROFINET IO industrial integration, sub-millisecond reject trigger output, and 24/7 production line thermal and reliability engineering.
The Engineering Problem
Machine Vision Is Among the Most Co-Dependent Engineering Domains in Industrial Automation
A 50µm defect on a part moving at 2m/s requires pixel pitch below 12.5µm, illumination pulse below 3ms for motion-free imaging, camera interface bandwidth above 4Gbps for the required frame rate, and inference pipeline latency below 8ms for the reject solenoid to fire before the part exits the inspection zone. Every one of these constraints fails in a system designed from component specification sheets without integrated application analysis — optics, sensor, interface, inference, and industrial I/O are co-dependent decisions.
A defect detector trained on 200 images under controlled lighting achieves 98% mAP in the laboratory and 65% accuracy on the production line — because the production line has three lighting variations between day and night shifts and a contamination pattern the training set never included. And a vision system driving a PROFINET IO reject output through a variable Python loop introduces 3–5ms jitter — enough to pass a defective part on a 600-parts-per-minute line.
Three structural drivers are expanding the market: Industry 4.0 platforms creating demand for vision systems that generate OPC-UA quality data and SPC-ready outputs alongside the reject trigger; NVIDIA Jetson, Hailo, and Movidius achieving edge AI compute density for complex segmentation models at production frame rates without cloud round-trips; and regulatory intensification under FDA 21 CFR Part 11, EU FMD, and FSMA demanding vision systems that generate compliant inspection records.
Industrial Quality Control & Optical Inspection Systems
AI-Powered Defect Detection & Deep Learning Vision
Food Sorting, Grading & Agricultural Vision
Robotic Bin Picking & Manipulation Vision
Dimensional Gauging & 3D Measurement Systems
Pharmaceutical Serialisation & Track-and-Trace
Security Analytics & Retail Vision
Medical & Scientific Imaging Electronics
Camera Interface & Image Acquisition
GigE Vision (1000BASE-T IEEE 802.3) with GenICam compliant device configuration and hardware trigger for deterministic acquisition timing. Camera Link Base/Medium/Full/Extended Full at up to 6.8Gbps aggregate bandwidth. CoaXPress CXP-6 and CXP-12 at 6.25Gbps and 12.5Gbps per lane. MIPI CSI-2 for embedded ARM and NVIDIA Jetson. FPGA-based frame grabber for deterministic acquisition without OS scheduling jitter. Multi-camera synchronisation with inter-camera phase alignment below 1µs for 3D and multi-view vision. Strobe synchronisation with sub-microsecond jitter for motion-freeze at high conveyor speeds.
Illumination Design & Control
Coaxial LED driver electronics for specular surface inspection. Darkfield illumination at oblique angle for surface scratch and pit detection — contrast derivation from BRDF physics of the specific surface. Structured light sinusoidal fringe projection electronics for 3D surface reconstruction. UV 365nm for fluorescent contamination detection. NIR 850nm for fill level through opaque containers. Multi-channel programmable LED driver with individual current control and thermal compensation. Strobe synchronisation achieving sub-5µs strobe-to-exposure alignment for 1m/s+ conveyor speeds. Illumination uniformity monitoring firmware for long-term consistency.
AI Inference Hardware & Pipeline
NVIDIA Jetson AGX Orin integration electronics with 275 TOPS DL inference and PCIe Gen4 for direct camera interface. Hailo-8 NPU at 26 TOPS inference in sub-3W for power-constrained embedded vision. Hailo-8L at 13 TOPS sub-1.5W for ultra-low-power vision. Intel Movidius Myriad X VPU for heterogeneous CPU/VPU/SHAVE core pipelines. FPGA-based neural network inference for deterministic sub-millisecond inference without GPU scheduling variance. TensorRT INT8 and FP16 model deployment with accuracy-throughput characterisation. OpenVINO for Intel edge platforms. End-to-end latency measurement from acquisition trigger to classification output.
Deep Learning Model Development
YOLOv8 and YOLOv9 object detection for multi-class defect detection at real-time throughput. Instance segmentation (Mask R-CNN, YOLOv8-seg) for simultaneous classification and defect boundary delineation. Semantic segmentation (DeepLabV3, SegFormer) for pixel-level defect mapping. Training dataset construction for production distribution — class imbalance management for 1000:1 good:defect ratio; synthetic defect augmentation via CycleGAN and diffusion-based generation; production variability characterisation. Model validation: precision, recall, F1, mAP on production-representative evaluation set. Model monitoring and drift detection for deployment quality assurance.
3D Vision & Structured Light
Laser triangulation profilometer interface electronics for inline 3D surface height measurement at line scan rates up to 10kHz. Structured light 3D reconstruction with phase-shifted sinusoidal fringe and GPU-accelerated phase unwrapping. ToF depth sensor interface for 3D bin picking at 30fps+. Stereo vision for dense point cloud generation. LiDAR sensor interface (Velodyne, Ouster, Livox, Hesai) for autonomous vehicle and AGV 3D perception. LiDAR-camera extrinsic calibration achieving sub-centimetre target point alignment. Point cloud processing pipeline for 3D object detection, classification, and 6D pose estimation.
Industrial Integration & Output Electronics
PROFINET IO device electronics with deterministic cycle time below 1ms for PLC reject trigger and status output. EtherNet/IP adapter for Allen-Bradley and Rockwell PLC integration. Sub-1ms NPN, PNP, and relay digital output from classification decision to output state change. Encoder input and conveyor position correlation for position-triggered acquisition and position-gated output. OPC-UA server with machine vision data object model for Industry 4.0 SCADA and MES integration. MQTT data streaming for quality analytics platforms. Serialisation data output (GS1 DataMatrix, QR) for pharmaceutical and food traceability. SPC data output with Cpk and Ppk calculation.
Optical System & Sensor Selection
Monochrome and colour CMOS and CCD sensor interface with EMVA 1288 characterisation for quantum efficiency, read noise, dynamic range, and dark current. Scientific CMOS (sCMOS) interface for sub-2 electron read noise in low-light and fluorescence inspection. Global shutter for motion artefact-free imaging. Telecentric lens mounting for distortion-free dimensional measurement. Colour vision with 3CCD or Bayer array for colour sorting and appearance inspection. SWIR InGaAs sensor interface for silicon wafer and agricultural moisture inspection. Multi-spectral and hyperspectral camera interface for material classification beyond visible spectrum.
Production Reliability & Thermal Management
MTBF prediction and component derating for 24/7 continuous operation. Thermal simulation for vision compute in sealed industrial enclosures. Heat pipe and vapour chamber thermal management for NVIDIA Jetson in fanless IP65 enclosures for washdown environments — validated at maximum ambient installation temperature at maximum inference load. Watchdog hardware for automatic system recovery from application failure without operator intervention. IEC 61000-4 series EMC immunity design for high-EMI manufacturing environments near VFDs and servo drives. Remote diagnostics and health monitoring for distributed vision system fleet management. Scheduled illumination lifetime and contamination monitoring firmware.
Chipsets & Platforms
Tested silicon and proven stacks — no experimental platform dependencies.
Industrial Quality Control, Defect Detection & Dimensional Gauging
Illumination Mode Selection From BRDF Physics: Why the Wrong Light Makes Defects Invisible
The gap between laboratory demonstration performance and production deployment is driven by factors the demonstration never includes: day-night illumination variation from skylights, batch-to-batch surface finish variation, conveyor belt contamination the algorithm classifies as a defect, and variable PLC reject timing. The BRDF of the surface determines which illumination mode maximises defect contrast: darkfield at 5–30° maximises scattered light from height features (scratch, pit, burr); coaxial reverses contrast on specular surfaces; structured light fringe distortion reveals surface height variation. Ankh derives the illumination architecture from BRDF physics of the specific surface and defect combination — not from the illumination products catalogue.
Robotic Vision, Agricultural Sorting & 3D Perception Electronics
LiDAR-Camera Sensor Fusion: Early vs. Late Fusion Architecture Trade-offs
Bin-picking 6D pose estimation must complete point cloud acquisition, part segmentation, CAD model matching, and grasp position computation — within the robot cycle time. LiDAR-camera extrinsic calibration achieves sub-centimetre 3D point alignment. Early fusion projects LiDAR features into the image domain for a single RGB-D network — better for small objects where camera spatial resolution captures detail the sparse LiDAR misses. Late fusion runs separate detection in image and point cloud domains before association — better for range-critical navigation where LiDAR metric depth corrects monocular depth estimation. Ankh selects the fusion architecture from the complementary sensing physics of the application.
Medical Imaging, Security Analytics & Retail Vision Electronics
GDPR Article 9 Biometric Data: The Legal Engineering Distinction That Matters
GDPR Article 9 special category processing — face recognition matching against enrolled identities — requires explicit legal basis, DPIA, and data subject rights management. On-device anonymous inference classifying age band and gender for aggregate audience statistics, without storing facial geometry, does not constitute biometric data processing under Article 9. A retail analytics system built on the latter delivers commercially valuable audience data within a demonstrably compliant privacy architecture. Ankh designs AI vision privacy architecture from the GDPR legal framework — not from a 'we don't store the images' claim that does not address the legal basis for processing.
Optical Architecture Designed From the BRDF Physics of the Specific Defect
A surface scratch on brushed aluminium requires darkfield illumination at a grazing angle — the same surface under coaxial illumination produces 10× less scratch contrast. The correct illumination mode is determined by the defect's BRDF: reflectance, depth, and orientation relative to surrounding texture. Ankh derives the illumination architecture from defect physics before the camera is selected and before the model is trained, so the AI receives the highest-contrast image optics can provide.
AI Model Training for the Production Distribution, Not the Laboratory Evaluation Set
A model achieving 98% mAP on a held-out set from the same distribution as training data provides no predictive evidence of production performance if illumination, orientation, or background texture differs on the line. Ankh designs the training data strategy from systematic production variability analysis — measuring shift-to-shift illumination variation, collecting defects across the full morphology range, and applying physics-informed augmentation — so the mAP on the production-representative evaluation set is the number that predicts deployment accuracy.
Reject Trigger Latency Measured End-to-End, Not Component-by-Component
The requirement is not inference latency — it is reject mechanism activation within X milliseconds of part crossing the camera axis, where X is set by conveyor speed and inspection-to-reject distance. Camera readout, interface transfer, preprocessing, inference, post-processing, PLC scan cycle, and solenoid response all contribute. Ankh measures end-to-end latency from camera trigger to reject output on production-representative hardware — not inference time on a development workstation.
24/7 Production Uptime Engineered From Component Selection Through Thermal Management
A vision system that performs during an 8-hour demonstration and develops illumination drift and thermal throttling after 6 weeks of continuous operation is not a production system. Ankh engineers production vision hardware with MTBF-appropriate component selection, thermal management validated at maximum installation ambient, watchdog hardware for automatic recovery, illumination intensity compensation for LED ageing, and remote health monitoring that alerts maintenance before a failure becomes a production stoppage.
Surface Defect Detection for an Automotive Stamping Press
Inline surface defect detection at 100fps, 4096×3072 resolution, end-to-end latency below 38ms at line speed — 8 defect classes including scratches, pits, and handling marks on steel panel surfaces.
Camera Link Full with FPGA frame grabber achieving deterministic 4ms acquisition latency at 100fps; 48-sector darkfield annular LED array for directional optimisation across panel orientation variation; YOLOv9 instance segmentation trained on 14,000 augmented examples across 8 defect classes; INT8 TensorRT on Jetson AGX Orin achieving 23ms end-to-end; PROFINET IO at 500µs cycle time; total trigger-to-NPN output 31ms on production hardware; OPC-UA quality data to MES.
- 99.2% detection rate for defects above 0.8mm at 100fps — validated on three-shift production data
- 31ms end-to-end latency from camera trigger to NPN reject output — within 38ms reject window at line speed
- 99.7% system uptime across three-shift continuous operation — Camera Link Full deterministic 4ms acquisition latency
- Replaced manual inspection team of 6 operators — OPC-UA quality data output integrated to MES for SPC and traceability
AI-Powered Fresh Produce Sorting System for a Food Technology Company
360° fresh produce inspection at 800mm/s conveyor speed, 20 items per second, with simultaneous colour, ripeness, blemish, and foreign object classification across four spectral bands and SPC output for grade consistency monitoring.
Four-station GigE Vision array (24 cameras) with multi-spectral LED illumination at 450nm, 550nm, 680nm, and NIR 850nm; YOLOv8 instance segmentation for blemish localisation alongside ripeness classification from multi-spectral features; 6× Jetson Orin Nano at 120fps processing 4 cameras each; encoder-correlated 12-zone air-blast reject at 20 items/second; 23ms end-to-end classification-to-reject output; SPC Cpk data output.
- 94.1% grade agreement versus trained human inspector at 3× throughput of prior manual grading process
- 28% reduction in good product misgraded as rejects — multi-spectral ripeness and blemish classification from 450/550/680nm and NIR 850nm
- 23ms end-to-end classification-to-reject output on production line — 6× Jetson Orin Nano modules processing 24 cameras at 120fps
- SPC Cpk data output for grade consistency monitoring — encoder-correlated 12-zone air-blast reject for individual item separation at 20 items/second
Building AI vision or machine vision electronics? Let's see what others miss.
Illumination architecture designed from the BRDF physics of the specific defect. AI model training for the production distribution rather than the laboratory evaluation set. Full-stack delivery from camera interface through inference hardware through PROFINET IO integration through 24/7 production thermal and reliability engineering.
Quote Your Vision Project