The Challenge
In the high-stakes world of manufacturing, maintaining quality control is paramount. For a leading global manufacturer, ensuring surface defect detection on their high-speed production line was a pressing challenge. Traditional inspection methods, which relied heavily on human intervention, were time-consuming and often inaccurate. This led to production bottlenecks and, ultimately, a delayed time-to-market. Furthermore, the manual process failed to achieve the desired defect detection accuracy, resulting in increased rework and waste.
The client required a robust solution that could keep pace with their production speed, provide real-time analysis, and achieve high accuracy in detecting surface irregularities. The goal was to automate the quality control process, reduce human error, and improve overall operational efficiency, ultimately ensuring only the highest quality products reach their customers.
The Solution
To address this challenge, the manufacturer integrated a cutting-edge Vision AI solution into their quality control processes. This state-of-the-art technology deployed advanced machine learning algorithms capable of real-time image processing and defect detection with outstanding precision.
The Vision AI system was designed to scan products as they traversed the high-speed line, identifying surface defects with a remarkable 99.4% accuracy rate. Leveraging AI models trained on thousands of defect images, the system could distinguish between acceptable variations and actual flaws requiring intervention. Not only did this approach drastically reduce the reliance on human inspection, but it also ensured consistent quality across all products.
Machine learning techniques allowed the system to continually improve its detection capabilities, learning from new defect images and updating its algorithms accordingly. Coupled with seamless integration into existing production line systems, the Vision AI solution facilitated swift deployment without significant disruption to the manufacturer's operations.
Key Results
The implementation of Vision AI technology resulted in transformative improvements across the client's manufacturing processes:
- Detection Accuracy: Achieved a 99.4% accuracy rate in defect detection, significantly reducing the number of faulty products reaching the end of the line.
- Increased Efficiency: Automating the quality control process reduced human inspection times, allowing operators to focus on higher-value tasks and enhancing overall production line efficiency.
- Cost Savings: Decreased the waste produced due to defects and the need for rework, optimising material usage and operational throughput.
- Faster Time-to-Market: Streamlined processes accelerated production timelines, ensuring products reached the market faster without compromising quality.
By deploying Vision AI, the manufacturer not only elevated their quality control measures but also solidified their position as an industry leader in deploying smart manufacturing technologies. This case exemplifies how innovative technology can be leveraged to overcome traditional manufacturing challenges, promoting enhanced product quality and operational excellence.
Technical Approach
Our solution was built on a multi-stage computer vision pipeline designed specifically for high-throughput industrial environments. The core inference engine used a custom-trained convolutional neural network (CNN) architecture based on EfficientDet, fine-tuned on a proprietary dataset of over 120,000 labelled surface images spanning twelve distinct defect categories — including micro-cracks, pitting, delamination, and contamination marks.
The hardware layer comprised industrial-grade GigE Vision cameras operating at 120 frames per second, paired with structured LED illumination arrays configured to reveal surface anomalies through directional light reflection. Edge inference was handled by NVIDIA Jetson AGX Orin modules mounted directly on the production line gantry, reducing latency to under 8 milliseconds per frame — well within the throughput requirements of the belt speed.
Key technology decisions included:
- Transfer learning from ImageNet-pretrained weights to accelerate model convergence on domain-specific defect imagery
- Data augmentation pipelines using Albumentations to synthesise defect variations and address class imbalance in training data
- ONNX Runtime for optimised inference deployment, enabling hardware-agnostic model portability
- MQTT-based messaging for near-real-time rejection gate actuation, ensuring flagged items were diverted within 40 milliseconds of classification
- MLflow for experiment tracking and model versioning, enabling systematic comparison of model iterations before production promotion
The system's architecture followed a modular design, separating the image acquisition layer, inference layer, and SCADA integration layer — making future upgrades to camera resolution or model architecture straightforward without requiring full redeployment.
Implementation Highlights
The project was delivered across four structured phases over a sixteen-week engagement.
Phase 1 — Data Collection and Labelling (Weeks 1–4): We embedded engineers on-site to capture a representative dataset across three production shifts, ensuring the training corpus reflected the full range of surface conditions, lighting variances, and product orientations encountered in real operation. A specialist labelling team annotated defect polygons using CVAT, applying strict inter-annotator agreement checks to keep labelling consistency above 97%.
Phase 2 — Model Development and Validation (Weeks 5–10): Multiple CNN architectures were benchmarked, including YOLOv8 and EfficientDet-D4. EfficientDet was selected for its superior balance between inference speed and detection precision at small defect scales. The model was validated against a held-out test set, achieving 99.4% accuracy and a false positive rate below 0.3% — a critical threshold to avoid unnecessary line stoppages.
Phase 3 — Edge Deployment and Integration (Weeks 11–14): The model was converted to ONNX format and optimised using TensorRT for deployment on Jetson hardware. Integration with the existing Siemens S7-1500 PLC was handled via OPC-UA, allowing the vision system to trigger rejection actuators autonomously. Extensive testing was conducted at full line speed before any live deployment.
Phase 4 — Hypercare and Monitoring (Weeks 15–16): A live monitoring dashboard was deployed using Grafana, surfacing real-time defect rates, model confidence distributions, and camera health metrics. A data drift detection mechanism was configured to flag when incoming image statistics deviated significantly from the training distribution, prompting scheduled model retraining.
The most significant challenge overcome was managing motion blur at high belt speeds. This was resolved by synchronising camera exposure timing with the belt encoder signal, achieving effective shutter speeds of 1/10,000th of a second.
Measurable Outcomes
The business impact of deploying Vision AI at scale went well beyond the headline accuracy figure. Prior to implementation, the client was experiencing quality escapes valued at approximately £2 million annually — products that passed manual inspection but were subsequently rejected by downstream customers or returned from the field.
Post-deployment results recorded over a six-month operational period demonstrated:
- Quality escapes reduced by 94%, bringing annual escape-related losses down from £2m to approximately £120,000
- False rejection rate of 0.3%, compared to an estimated 2.1% false rejection rate under manual inspection — recovering significant yield value
- Inspector headcount redeployment: four full-time quality inspectors were transitioned to process improvement roles, representing an annual labour saving of approximately £140,000
- Mean time to defect identification fell from 4.2 minutes (manual review cycle) to under 10 seconds (automated flagging and alert dispatch)
- Line utilisation increased by 6.8% as a direct result of reduced unplanned stoppages caused by escaped defects triggering downstream rejections
When combined with reduced rework costs, material savings, and customer chargeback avoidance, the client's internal ROI calculation showed full payback of the project investment within eleven months.
Why This Approach Worked
The success of this deployment was rooted in a combination of rigorous data engineering and close operational co-design with the client's production team — neither element alone would have achieved the outcome.
Many vision AI projects fail in production not because the model is technically weak, but because the hardware setup produces images that are inconsistent with training conditions. By designing the illumination and camera configuration before any data collection took place, we ensured that the training dataset authentically represented what the deployed system would see under all shift conditions and seasonal lighting variations within the facility.
The choice of edge inference over cloud-based processing was equally deliberate. Latency introduced by network hops — even on a low-latency LAN — would have been incompatible with the rejection gate timing requirements. Running inference locally on Jetson hardware eliminated this constraint entirely.
Finally, the MLflow-based experiment tracking and staged model promotion workflow meant that future model improvements could be validated in shadow mode against live production data before any cut-over. This gave the client's engineering team the confidence to own ongoing model maintenance without dependency on external consultants for routine updates.
Lessons Learned
Several insights from this engagement have since informed how we approach industrial vision AI projects more broadly.
Lighting design is as important as model design. Early prototype sessions with off-the-shelf ring lights produced inconsistent results on curved product surfaces. Switching to a bespoke four-quadrant directional illumination rig — where each quadrant could be individually controlled — revealed defect categories that had been systematically missed. Investing time in optics before writing a single line of model code proved to be the highest-leverage decision in the project.
Class imbalance must be addressed at the data level, not solely through loss function weighting. The rarest defect category — sub-millimetre delamination — appeared in fewer than 0.8% of training samples. We found that synthetic augmentation using domain-randomised overlays, combined with targeted re-sampling during training, outperformed weighted cross-entropy adjustments alone for this class.
Operator trust is a deployment risk. The production team was initially sceptical of automated rejection decisions, particularly on borderline cases. Incorporating a review queue — where low-confidence classifications were surfaced to a human reviewer rather than rejected outright — was not technically necessary given the model's performance, but it was essential for building operational confidence during the first weeks of live running. Within three weeks, the review queue was seeing fewer than five borderline cases per shift, and operators were consistently validating the automated decisions, which accelerated full trust in the system.
Speak with our Computer Vision team at Adyantrix to find out how we can support your next project.
Work with Adyantrix
If you are looking to tackle a similar challenge, Adyantrix has the expertise to help across the full project lifecycle. Our computer vision practice covers image classification, object detection, and inspection AI. Our AI & machine learning practice covers ML model development, MLOps, and intelligent automation. Our ML model development practice covers supervised, unsupervised, and deep learning models. Our data analytics practice covers BI reporting and self-serve analytics platforms. Get in touch to discuss your requirements — no commitment required.
