AI Quality Control: Visual Inspection on AWS

Market Context: AI Quality Control in German Manufacturing

Manual visual inspection is one of the last major cost drivers in otherwise highly automated production lines. A trained inspector reliably detects surface defects, dimensional deviations and assembly errors — but only under adequate lighting, without fatigue effects and at low throughput speeds. In modern high-speed lines, human inspection hits physical limits.

According to the Bitkom study 2024, 42% of German manufacturing companies are already using AI technologies in production or are in the pilot phase. Visual inspection — automated image analysis for defect detection — leads the ranking of prioritised use cases. The reasons are straightforward: defect rates drop by 40–80%, inspection costs by 30–60%, and cycle time decreases measurably.

At the same time, the EU AI Act introduces new compliance requirements for AI systems in production. Manufacturing companies face the task of implementing AI quality control not only technically, but also documenting and operating it in a regulatorily sound manner.

Terminology: The AI Vocabulary of Visual Inspection

A shared vocabulary is essential for productive discussions between technology and compliance teams. The following definitions form the basis for all sections of this article.

Computer Vision

A subfield of artificial intelligence that develops algorithms for interpreting digital images and videos. In manufacturing, computer vision encompasses defect classification, object detection, dimensional measurement and surface analysis. Systems learn from labelled image examples which patterns are to be assessed as defective.

Convolutional Neural Network (CNN)

A neural network architecture specifically developed for image processing. CNNs recognise spatial patterns through hierarchical filters: early layers detect edges and textures, deeper layers combine these into complex features such as cracks, scratches or shape deviations. ResNet, EfficientNet and MobileNet are common CNN architectures in quality control.

Transfer Learning

A technique in which a model pre-trained on large image datasets (e.g. ImageNet) is used as the starting point for a specific task. The pre-trained layers already know general image features — the new model only needs to learn the task-specific differences. Transfer learning typically reduces training data requirements by 80–95% compared to training from scratch.

Inference

The process by which a trained model makes a prediction on new input data (production images). A distinction must be made between cloud inference (the image is sent to the AWS cloud, evaluated there, and the result returned) and edge inference (the model runs directly on a device in the factory). Edge inference is mandatory for real-time inspection.

Confidence Score

A number between 0 and 1 indicating how certain the model is in its prediction. A confidence score of 0.95 for "defective" means the model has detected a defect with 95% probability. The threshold above which a part is classified as scrap is a critical configuration parameter: too low leads to false positives (good parts are rejected), too high to false negatives (defective parts pass the inspection).

The ML Pipeline: From Raw Data to a Production-Ready Model

An AI quality inspection system is not created by switching on a service, but through a structured pipeline that connects data, model and infrastructure. The following steps apply to all AWS-based visual inspection setups.

1. Camera installation and lighting optimisation: Image quality is the most important prerequisite for high model performance. Industrial cameras (GigE Vision or USB3 Vision) are triggered via PLC trigger signals. Coaxial lighting, ring light or fringe projection lighting are selected depending on the type of defect. This step determines 60% of the subsequent detection accuracy.
2. Data collection and organisation: Images are systematically collected and organised into folders by class (Normal, Scratch, Crack, Incorrectly assembled, etc.). Amazon Kinesis Video Streams handles continuous video capture and enables the extraction of single frames for labelling. Goal: at least 100–300 images per defect class.
3. Labelling with Amazon SageMaker Ground Truth: Quality assurance specialists label images via a web interface. SageMaker Ground Truth supports bounding box annotations, segmentation masks and image classification. Auto-labelling accelerates the process once a first model has been trained.
4. Model training: Depending on data volume and requirements: Amazon Lookout for Vision (no-code) or SageMaker Training Jobs with custom algorithms. SageMaker Experiments tracks all training runs, hyperparameters and metrics for reproducibility and documentation in line with the EU AI Act.
5. Model validation: The model is evaluated on a separate test dataset: precision, recall, F1-score and confusion matrix. For safety-critical components: recall (sensitivity) for the defect class must be close to 100% — false negatives are generally more dangerous than false positives.
6. Deployment and integration: The validated model is deployed as a SageMaker Endpoint (cloud) or as a Greengrass component (edge). Integration into the MES is handled via AWS Lambda and REST/MQTT interfaces. Results are persisted in Amazon S3 and Amazon DynamoDB.
7. Monitoring and retraining: Amazon SageMaker Model Monitor tracks data drift and concept drift — when production reality changes (new supplier, changed manufacturing parameters), model accuracy declines. Automated retraining triggering ensures the model stays current.

Amazon Lookout for Vision: The No-Code Entry Point

For manufacturing companies without a data science team, Amazon Lookout for Vision is the fastest path to AI quality control. The service requires no programming knowledge and is optimised precisely for this use case: distinguishing between normal and defective products based on images.

The process is straightforward: upload images of good parts and defective parts to Amazon S3, create a model version in the Lookout for Vision project, and start training. The service automatically selects the appropriate architecture, optimises hyperparameters and returns a model with confidence scores. Typical training time: 30–90 minutes.

Lookout for Vision is ideal for the pilot and for use cases with clearly defined defect classes. Once multi-stage classifications, dimensional measurements or anomaly-based detection without defect labels are needed, switching to SageMaker Custom Models makes sense.

SageMaker Custom Models: A Tailored Computer Vision Pipeline

When requirements go beyond a binary good/defective verdict, Amazon SageMaker provides full control over the computer vision pipeline. SageMaker supports PyTorch, TensorFlow and MXNet — the three dominant frameworks for industrial computer vision.

A typical custom model workflow on SageMaker begins with a pre-trained base model (e.g. ResNet-50 from the SageMaker JumpStart model catalogue) that is adapted to the specific dataset via transfer learning. Hyperparameter optimisation is handled by SageMaker Automatic Model Tuning, which runs parallel training jobs on managed instances.

For anomaly detection without defect labels — when sufficient defective images are not available — autoencoder architectures trained on good parts and detecting deviations as anomalies are well-suited. AWS provides reference implementations for this approach in SageMaker Notebooks.

SageMaker Pipelines formalises the entire ML workflow as repeatable, versioned code: data preprocessing, training, validation and deployment are defined as pipeline steps and can be triggered automatically by code changes or new data. This reproducibility is a core requirement of EU AI Act documentation obligations.

Edge Inference: Real-Time Inspection Directly at the Line

Cloud inference is too slow for most visual inspection use cases. Sending an image to the cloud, evaluating it there and receiving the result back typically takes 200–500 ms — but on a line running 60 parts per minute, only about 1 second is available for the inspection decision. Edge inference solves this problem by running the model directly on a computer in the factory.

AWS IoT Greengrass V2 enables the deployment of SageMaker models to edge devices: industrial PCs, NVIDIA Jetson modules or specific AI accelerators. Greengrass manages the model lifecycle — updates are rolled out via the cloud without manual intervention at the line.

Runtime optimisation is achieved through model compression: NVIDIA TensorRT or AWS Neuron compile the model for the specific hardware and reduce inference latency by 3–10x. A ResNet-50 that requires 180 ms in cloud inference achieves under 15 ms on an NVIDIA Jetson AGX Orin with TensorRT optimisation.

The communication architecture between edge and cloud is asymmetric: inference results (classification, confidence score, image reference) are sent to AWS IoT Core in real time. The original images — especially from rejected parts — are uploaded asynchronously to Amazon S3 for retraining and audit purposes. Amazon Kinesis Video Streams can be used for continuous video recording when gapless documentation is required.

ROI Calculation: What AI Quality Control Concretely Delivers

The economic assessment of an AI quality inspection system follows a structured model that balances cost savings against investment costs. For a mid-sized manufacturing company with a single inspection line, the following levers are relevant:

The break-even in this example is under 6 months. Critical to the accuracy of the ROI calculation is capturing the true cost of defects: scrap costs, rework costs, warranty costs and reputational costs from customer complaints. Many companies underestimate the total cost of a high defect rate because it is spread across multiple cost centres.

EU AI Act and AI Quality Assurance: What Applies from 2026?

The EU AI Act has been in force since August 2024 and is becoming applicable in stages. For manufacturing companies using or planning AI quality control, concrete obligations arise — particularly when the systems make safety-relevant decisions.

Risk Classification under the EU AI Act

AI systems used in the context of the Machinery Regulation (EU) 2023/1230 that make decisions affecting product safety typically fall into the high-risk category (Annex III, No. 6b). This means:

• Risk management system for the AI system throughout the entire lifecycle
• Data governance and data quality assurance (training data, validation data)
• Technical documentation in accordance with Annex IV of the EU AI Act
• Automatic logging of all system decisions
• Transparency and user information: operators must know they are working with an AI system
• Human oversight: mechanism to intervene and override the system
• Robustness and accuracy: documentation of metrics and thresholds
• Conformity assessment before deployment and regular review

AWS supports these requirements through the native auditability of the SageMaker stack: SageMaker Experiments documents all training runs, SageMaker Model Registry versions models with metadata, and Amazon CloudWatch Logs records all inference decisions with timestamp and confidence score.

GDPR Aspects for Image Data

If camera images capture persons (e.g. operators at the line), GDPR requirements apply: purpose limitation, data minimisation, technical and organisational measures. In practice, the camera field of view is aligned so that only the product is captured. Where this is not possible, privacy masking or edge-side blurring algorithms must be applied before images leave the factory.

Storm Reply Perspective: What Works in Practice

Storm Reply supports manufacturing companies in implementing AI quality control on AWS — from the initial feasibility study to the production-stable system. Recurring insights from these projects include:

Data quality beats model architecture. The most common reason for unsatisfactory detection rates is not an overly simple model, but too few or poorly lit training images. In 80% of cases, improving camera setup and data collection delivers more than switching model architecture.

Start with Lookout for Vision, scale with SageMaker. The pragmatic path for new customers: Amazon Lookout for Vision for the first pilot to quickly demonstrate results and build internal acceptance. Once the system is live and more requirements emerge, we selectively migrate to SageMaker Custom Models — only where the added value justifies the additional complexity.

Build EU AI Act compliance in from the start. Projects that address regulatory requirements from the beginning in the architecture (logging, documentation, human-in-the-loop) are substantially cheaper to qualify later than systems where these aspects must be retrofitted.

Edge inference is often the critical path. Camera integration, network segregation (OT/IT) and deployment on edge hardware are technically more demanding than model training itself. We use standardised Greengrass deployment pipelines that make this step reproducible and maintainable.

Advantages and Challenges at a Glance

Frequently Asked Questions about AI Quality Control on AWS

What is the difference between Amazon Lookout for Vision and a SageMaker Custom Model?

Amazon Lookout for Vision is a no-code service that works with a few hundred example images and is optimised for binary good/defect classification. SageMaker Custom Models enable bespoke CNN architectures, multi-label classification and segmentation — but require datasets in the thousands range and data science expertise.

How many images do I need to start an AI quality inspection system?

Amazon Lookout for Vision requires at least 20 images per class (good and defective), practically 100–300 images per defect class are recommended. SageMaker models achieve robust results from around 1,000 images per class. Transfer learning significantly reduces data requirements because pre-trained models already know general image features.

What latency does edge inference achieve with AWS IoT Greengrass?

With optimised models (NVIDIA TensorRT, AWS Neuron), inference on an edge device typically achieves 10–50 ms per image. This enables inspection at line speeds of up to 100 parts per minute. Cloud communication for audit logging and model updates runs asynchronously and does not block the production flow.

How is AI quality control classified under the EU AI Act?

Automated quality control systems in manufacturing typically fall into the high-risk category under the EU AI Act when they make decisions that directly affect product safety (Machinery Regulation). This requires: a risk management system, technical documentation, data governance, human oversight and conformity assessment before deployment.

How long does a visual inspection pilot project with AWS take?

A focused pilot on a single inspection station is achievable in 6–10 weeks: weeks 1–2 camera installation and data collection, weeks 3–5 model training and validation, weeks 6–8 Greengrass deployment and production integration, weeks 9–10 qualification and go-live. Full rollout to multiple lines takes a further 3–6 months.

Sources and Further Reading

• Bitkom e.V.: AI in Industry — Status and Prospects 2024, Berlin 2024
• European Commission: Regulation (EU) 2024/1689 — EU AI Act, Official Journal of the EU, August 2024
• European Commission: Regulation (EU) 2023/1230 — Machinery Regulation
• AWS Documentation: Amazon Lookout for Vision Developer Guide, AWS Inc.
• AWS Documentation: Amazon SageMaker Developer Guide — Computer Vision, AWS Inc.
• AWS Documentation: AWS IoT Greengrass V2 — Edge ML Inference, AWS Inc.
• VDI/VDE 2632: Industrial Machine Vision — Standards and Guidelines