●[PROJECT NAME: Real-Time Defect Detection for Manufacturing Lines]
Replacing slow manual surface inspections with an edge-optimized YOLOv8 pipeline that spots micro-defects in 14 milliseconds.
When manufacturing items glide past a high-speed camera, human eyes get tired after 20 minutes. We trained a model to spot hairline fractures, scratches, and misalignment in real time, alerting the conveyor arm to discard broken items before packaging.
The factory floor operated on manual spot-checks where inspectors examined parts under high-glare LED bars.
Inspectors caught roughly 82% of surface flaws at shift start, but fatigue reduced consistency by mid-shift. The hardware partner needed an automated edge vision system that could run directly inside the inspection housing without relying on an external cloud connection.
Crucially, lighting conditions fluctuated by 18% throughout the day due to skylight variation, confounding standard OpenCV edge-detection filters and causing unacceptable false rejects.
“If an inspection algorithm takes more than 20 milliseconds per frame, the conveyor belt has already moved the part past the pneumatic rejection arm.”
Batch size 1 on NVIDIA Jetson Orin Nano
Across 4 defect classes on custom industrial dataset
Curated, augmented, and contrast-balanced
Compared against the baseline heuristic thresholding
Architectural choices and what was explicitly rejected.
Selected YOLOv8-Nano quantized with TensorRT over larger ResNet backbones
Heavy transformer vision models offered marginal accuracy gains (+1.2% mAP) at the cost of 4x higher latency and double the thermal footprint on edge hardware. Quantizing YOLOv8n to INT8 brought inference latency under 15ms while keeping accuracy above 94%.
Rejected cloud API roundtrips (too slow, unreliable local network) and large multi-stage mask-RCNN architectures.
Synthetic shadow and specular glare augmentation during training
Rather than installing expensive polarizing filter glass on the factory ceiling, we generated randomized specular glare patches and Gaussian contrast shifts across 3,000 baseline images to make the model invariant to daylight shifts.
Rejected relying solely on factory-captured validation images.
High-framerate GigE camera capture feeds an in-memory ring buffer, processed via OpenCV CUDA bindings into a TensorRT execution engine running on edge hardware.
Optical Capture & Region Normalization
Industrial camera streams 60fps at 1080p. Frame is cropped to the conveyor ROI and normalized.
TensorRT INT8 Inference Engine
Quantized weights evaluate bounding boxes and surface classification masks in under 15ms.
GPIO Signal & Telemetry Egress
If confidence exceeds 0.88, a hardware trigger actuates the reject valve; metadata is batched to SQLite.
The automated pipeline ran continuously for [TEST DURATION: 3-WEEK PILOT], reducing manual inspection overhead and catching micro-cracks before shipping.
- 01.Edge hardware thermals matter just as much as validation loss.
- 02.Data augmentation targeting real environmental noise (lighting, dust) outperforms raw model size every time.
- 03.Clean fallback logic is required when a camera lens collects surface dust.
“Computer vision in the real world is 80% data curation and deployment constraints, and 20% model selection.”