OptiInspect-Edge addresses the extreme quality demands of electronics manufacturing lines. Integrating a high-speed global-shutter CMOS camera, custom telecentric coaxial strobe illumination, and an embedded GPU inference engine, the system performs automated optical inspection (AOI) on moving printed circuit board (PCB) conveyor lines. It classifies micron-level solder bridges, tombstoned passives, missing components, and polarity errors in real time at 120 frames per second.
High-density surface mount technology (SMT) with 0201-metric components (0.3mm x 0.15mm) is impossible to inspect manually with human eyes. Existing commercial Automated Optical Inspection (AOI) machines cost upwards of $150,000, occupy massive shopfloor footprints, and rely on rigid rule-based vision algorithms that generate high false-alarm rates (> 15%), forcing expensive manual reinspection bottlenecks.
Design a compact industrial machine vision optical assembly with zero perspective distortion using a telecentric lens.
Construct a multi-angle RGB LED ring strobe illuminator that highlights solder meniscus specular reflections while eliminating shadows.
Develop a lightweight convolutional attention network (Tiny-PCBNet) detecting 8 distinct defect classes in under 7.5 milliseconds per frame.
Deploy on an industrial embedded edge platform providing real-time pneumatically actuated bad-board reject triggering over industrial IO.
The imaging and classification pipeline:
- Camera & Optics: Sony Pregius IMX392 global shutter CMOS sensor (3.2 MP, 120 FPS, USB3 Vision) paired with a 0.5x bi-telecentric lens (depth of field ±4mm).
- Illumination: Triple-ring RGB coaxial LED array with microsecond constant-current strobe controller synchronized to camera exposure.
- Edge Processing: NVIDIA Jetson AGX Orin industrial edge computer running TensorRT with INT8 quantized neural models.
- Machine Control: Optoisolated digital I/O interface connecting directly to Siemens / Omron conveyor line PLCs.
- Dataset: Curated a dataset of 45,000 high-resolution cropped SMT component images containing real manufacturing anomalies.
- Defect Taxonomy: Solder Bridging, Tombstoning, Insufficient Solder, Component Shift, Lifted Lead, Wrong Polarity, Solder Balling, and Missing Part.
- Benchmark: Evaluated against conventional morphological template matching and full-sized server-grade ResNet architectures.
As circuit boards pass beneath the optical head at speeds up to 40 meters per minute, the system captures non-blurred frames. Solder joints are highlighted in red, green, and blue based on light reflection angle. Defect classifications are displayed on the operator inspection monitor with visual overlays, while an ejection air-knife blows rejected boards into the quarantine chute.
Testing analyzed defect recall rates across varying conveyor line velocities (0 to 0.8 m/s), evaluated illumination stability against ambient factory ceiling lighting, and measured thermal drift on the edge GPU under 72 hours of continuous 120-FPS processing.
- Defect Detection Accuracy: 99.4% recall across all 8 defect categories with 0.4% false-positive rate.
- Processing Frame Rate: 124 FPS sustained throughput at full 3.2 MP resolution.
- Spatial Resolution: 14.8 microns per pixel across a 30mm x 22mm inspection field of view.
- Pneumatic Reject Response Time: 12.2 milliseconds from image capture to digital output pulse.
- Hardware: Sony IMX392 Camera, Telecentric Lens, Strobe Driver, NVIDIA Jetson AGX Orin, Industrial 24V I/O Board.
- Software: C++20, OpenCV, CUDA, TensorRT, PyTorch, Qt6 Industrial UI, Modbus TCP / Industrial Ethernet.
Annotated open benchmark dataset of 45,000 industrial SMT solder and component defects.
TensorRT-optimized deep learning inference pipeline with zero-copy memory transfers.
Optical CAD mechanical mount designs with vibration-damped linear adjustments.
Formal industrial vision deployment and photometric strobe calibration guideline.
Automated electronics assembly lines, automotive ECU manufacturing quality assurance, aerospace avionics PCB verification, and medical implant electronics inspection.
High-intensity LED strobes can present optical hazards. The inspection bay is fully enclosed with interlocked optical baffling to prevent operator exposure to high-luminance light pulses.
A compact, low-cost embedded machine vision inspection station delivering 120-FPS high-precision electronics defect classification matching the reliability of commercial multi-hundred-thousand-dollar AOI machines.
Incorporating multi-spectral infrared imaging to detect hidden micro-voids underneath BGA (Ball Grid Array) component solder balls without requiring destructive X-ray equipment.
Micron-level automated optical inspection classifying electronics manufacturing defects at 120 frames per second.
An embedded industrial computer vision system combining telecentric optics and TensorRT-optimized deep neural networks to inspect surface-mount electronics assemblies at 120 FPS with 15-micron defect resolution.
