Use case

Automating quality control with computer vision

How industrial SMBs deploy computer vision to automate quality control on their production lines. Complete methodology, technologies and measurable ROI.

8 min read
Vision IAQualitéIndustrieAutomatisationROI
⚡ The case in 30 seconds

Computer vision is revolutionizing industrial quality control

Visual quality control remains the weak link of many production lines: operator fatigue, subjectivity of judgments and the impossibility of inspecting 100% of parts. Computer vision, driven by advances in deep learning, now makes it possible to inspect each part in under 200 milliseconds with a defect detection rate above 99%. And costs have been cut by a factor of 5 in two years, making the technology accessible to industrial SMBs.

This use case details the deployment of an AI vision quality control system at an industrial subcontractor with 80 employees in the Hauts-de-France region, from the initial audit to the results measured after 6 months of operation.

The problem: costly and fallible human quality control

Industrial SMBs face growing quality requirements from their clients, while having to keep their production costs under control. Three recurring pain points stand out.

👁️

Visual fatigue and degraded detection rate

A visual quality control operator sees their detection rate drop from 95% to 70% after 2 hours of continuous inspection. Over an 8-hour shift, the average rate falls to 80%. This inevitable fatigue lets defects through that generate customer returns, contractual penalties and damage to brand image.

📉

Insufficient sampling-based control

For lack of capacity, most SMBs inspect between 5 and 20% of their production. Defective batches slip through the statistical net. A plastic parts manufacturer was delivering on average 3 non-compliant batches per quarter despite a 10% sampling control, representing a non-quality cost of 45,000 euros per year in returns and rework.

💶

Rising cost of quality labor

Recruiting and training qualified quality operators is becoming increasingly difficult and costly. The full cost of a human inspection station (salary, charges, training, management) reaches 45,000 to 55,000 euros per year. With the average turnover of 15% in these roles, the investment in training is constant.

The solution: an AI vision system integrated into the production line

Our approach deploys a computer vision system directly on the production line, capable of inspecting 100% of parts in real time.

📸

High-resolution image acquisition

High-resolution industrial cameras (5 to 12 megapixels) capture each part from one or several angles with controlled LED lighting. The system processes up to 30 parts per second, compatible with standard production rates. Structured lighting (dome, backlight, grazing light) is adapted to the type of defect being sought in order to maximize contrast.

🧠

Defect detection by deep learning

A convolutional neural network (CNN) trained on images of your production learns to distinguish compliant parts from defective ones. The model detects cracks, scratches, deformations, stains and dimensional deviations with an accuracy of 99.2%. It continuously improves thanks to operator corrections on borderline cases.

Automatic sorting and full traceability

The system commands an ejection mechanism (cylinder, air blast, diverter) that automatically separates non-compliant parts. Each inspection is recorded with the image, the verdict, the model's confidence and the timestamp. This complete traceability log satisfies the requirements of the ISO 9001 and IATF 16949 certifications and enables a statistical analysis of quality trends.

Step-by-step implementation

1

Quality audit and image collection (week 1-4)

Analyze your quality history: types of defects, frequency, associated costs. Identify the pilot inspection station (the one with the highest business impact). Install a temporary acquisition system to collect 2 to 4 weeks of images in real production conditions. Annotate the images with the quality operators to build the training set: a minimum of 300 images per defect type, plus 1,000 images of compliant parts.

2

Model training and validation (week 5-8)

Train the detection model on your dataset with data augmentation. Validate on a representative test set (20% of images held out). Minimum target: 98% detection of critical defects with less than 2% false positives. Iterate on lighting and camera positioning if performance is insufficient. Have the results validated by the quality team on 500 parts inspected in parallel (human + machine).

3

Line integration, ramp-up to production (week 9-14)

Install the final hardware (industrial cameras, lighting, processing PC, ejection mechanism) during a scheduled maintenance window. Run the system in "observation" mode for 2 weeks (alert without ejection) to validate in real conditions. Switch to automatic mode with human validation of ejections for 2 additional weeks. Move to full autonomy once the false positive rate is stabilized below 1.5%. Train the operators to supervise the system and handle alerts.

Results observed after 6 months

Detection rate
99.2% (vs 80% human)
Parts inspected
100% (vs 10% before)
Non-quality cost
-72% over 6 months
ROI
Reached in 11 months

Frequently asked questions

How many images do I need to train a defect detection model?

It depends on the complexity of the defects. For visually obvious defects (crack, hole, deformation), 200 to 500 annotated images are enough for a first functional model. For subtle defects (color variation, micro-scratch), allow 1,000 to 3,000 images. Data augmentation techniques (rotation, zoom, brightness change) let you virtually multiply your dataset. In practice, we recommend collecting 2 to 4 weeks of production to build a representative dataset.

Does the system work in a harsh industrial environment?

Yes, but it requires particular attention to lighting and camera positioning. Vibrations, dust, and temperature and light variations can affect image quality. We recommend IP67 industrial cameras with controlled LED lighting and suitable protective enclosures. Regular calibration (monthly) ensures stable performance. Modern solutions include compensation algorithms that tolerate 15 to 20% variation in lighting conditions.

Can computer vision be integrated into an existing production line?

Absolutely, it is even the most common scenario. Installation is generally done without stopping production: mounting the cameras and lighting during scheduled maintenance (4 to 8 hours), then a calibration phase in parallel with existing production. Integration with the sorting/ejection system is done via standard signals (PLC, OPC-UA). The vision system runs in parallel with human inspection for 2 to 4 weeks before taking over.

What budget should I plan for a first automated inspection station?

For an inspection station with one camera, allow between 15,000 and 40,000 euros all-in: hardware (industrial camera, lighting, processing PC) for 5,000 to 12,000 euros, model development and training for 8,000 to 20,000 euros, and mechanical/electrical integration for 2,000 to 8,000 euros. Operating costs are low: 100 to 300 euros per month for maintenance and model updates. ROI is generally reached in 8 to 14 months.

Industrial vision technologies

YOLOv8 / YOLOv9

Real-time object detection

The most popular detection architecture in industrial vision. Capable of detecting and locating defects in under 20 ms on a standard GPU. The nano model (YOLOv8n) even runs on an industrial PC without a dedicated GPU for simple cases. Open source and actively maintained.

Anomalib (Intel)

Unsupervised anomaly detection

Open source library specialized in visual anomaly detection. Major advantage: it only requires images of compliant parts for training (no need to collect rare defects). Ideal when defects are varied and unpredictable. Accuracy of 95 to 98% depending on the case.

Pricing

Industrial camera 800-3 000 €
Processing PC 2 000-5 000 €
Model development 8 000-20 000 €

Comparison

Criterion AI vision (deep learning) Classic vision (rules) Human inspection
Detection rate 99%+ 90-95% 70-85%
Coverage 100% 100% 5-20%
Adaptability to new defects Retraining Reprogramming Immediate
Annual cost (1 station) 3-5 K€ 5-10 K€ 45-55 K€

Related articles