This article guides manufacturing companies in implementing Computer Vision for quality control, with the goal of moving toward the Zero Defect model. It begins with the current context (where the production mix is growing but inspection often remains manual and sample-based) to explain how computer vision enables 100% automated inspection that is more standardized, continuous, and data-driven.
It analyzes the two main approaches (classical Machine Vision and Deep Learning) and the value of a hybrid strategy, along with the fundamental components for building a reliable industrial dataset: scene, lighting, labeling, and model lifecycle.
The article then moves on to reference standards (ISO 9001, EMVA 1288, OPC UA, ISO/IEC 42001) and updated ISTAT and Eurostat 2025 data, which reveal that adoption remains selective and thus presents a concrete competitive opportunity.
The core of the guide is a 6-step roadmap, from CTQ selection to industrialization and MLOps, accompanied by the 5 most common mistakes to avoid. The conclusion reiterates that Zero Defect is not an algorithm, but an integrated system of measurement, decision-making, and continuous improvement.
Computer Vision enables manufacturing companies to transition from manual and sample-based inspections to a 100% automated visual inspection model, reducing defects and scrap in a measurable and repeatable way. In this guide, you’ll find up-to-date data, reference standards, and a concrete roadmap for implementing it with industrial rigor.
“Zero-defect” quality is not just an ambitious goal: it is the result of a system that measures, decides, and improves in a structured way. The technology truly works when it is designed as a process system, supported by reliable data, and anchored in standards for quality, measurement, integration, and AI governance.
💡 Want to understand where your company stands on the path to the Smart Factory? Discover Bonfiglioli Consulting’s Operations 2025 Benchmarking Study →
In the factory, quality control often faces a paradox: production mix and variability are increasing, but inspection capabilities remain tied to manual sampling and checks. Machine vision resolves this imbalance by making the process more standardized—same scene, same criteria, less operator subjectivity—more continuous, with inspections up to 100% in-line where necessary, and more data-driven, because every image generates a traceable result and a corrective or preventive action.
Robustness increases further when different modalities are combined: 2D for contrast, texture, and color; 3D for shape and surface defects. An often-overlooked factor is the need to account for the ongoing maintenance of data and models: this is not a failure, but a natural characteristic of any industrial AI system.
When discussing vision for quality control, it is useful to distinguish between two main approaches. Classic Machine Vision is based on predefined rules and thresholds: it is deterministic, fast, and easy to audit, but requires stable conditions and well-defined defects. Computer Vision with Deep Learning, on the other hand, is better suited when finishes, textures, reflections, batches, and process conditions change, or when defects are numerous, rare, or heterogeneous.
In practice, the most robust projects are often hybrid: rules for simple constraints and deterministic checks, and AI for classification and anomaly detection where variability makes fixed-threshold logic unreliable. The choice of approach is not ideological but depends on the type of defect, the stability of the scene, and the required level of traceability.
This is where almost everything is decided. An automated “factory-floor” visual inspection system requires treating the image as process data, not as a simple photo.
Even before training, the scene must be made repeatable: position, distance, exposure times, vibrations, reflections, and physical barriers. Lighting is often the variable that “makes or breaks” data quality—and the most common mistake is to postpone its standardization until later.
If using AI, labels must be consistent: defined defect taxonomy, clear rules for borderline cases, “golden set” samples for regression testing. The model must be updated periodically to account for new cases and process variability: this is normal; it’s part of the lifecycle.
Machine vision systems are not “set & forget.” They require support for development and maintenance and, often, new roles and internal responsibilities. Devices (cameras, edge computing) are subject to rapid obsolescence: it is advisable to design from the outset with replaceability and standardization in mind.
To bring Computer Vision into quality control with industrial rigor, it is useful to rely on recognized standards at multiple levels.
ISO 9001 is the most widely used reference for establishing and improving a quality management system: processes, documentation, and continuous improvement.
EMVA 1288 is the standard for measuring and presenting the specifications of machine vision sensors and cameras in a comparable manner: useful for technical selection, specifications, and system validation.
GenICam offers a plug-and-play interface to manage cameras and devices through a common interface, reducing complexity and vendor lock-in. OPC UA for Machine Vision, on the other hand, allows inspection systems to be integrated with production control and IT systems, enabling vertical and horizontal integration of quality data throughout the supply chain.
ISO/IEC 42001 provides the standard for establishing an AI management system with roles, controls, and continuous improvement. ISO/IEC 23894 offers guidance on managing risks specific to AI, while the NIST AI RMF 1.0 is the international reference framework for managing risks and the reliability of AI systems throughout their entire lifecycle.
When vision is integrated into measurement and decision-making processes, it is important to consider accuracy and repeatability, just as with any measurement system. On the prevention side, tools such as FMEA/FMECA support the prioritization of controls on CTQs (Critical To Quality).
Current figures send a clear message: in manufacturing, quality is already one of the most mature AI use cases, but the adoption of computer vision on an “all-business” basis is still selective. The gap between those who have already industrialized these solutions and those still in the PoC phase represents a concrete competitive opportunity.
In global manufacturing, the Google Cloud report “AI acceleration among manufacturers” indicates that 39.1% of companies already use AI for quality inspection and 35.1% for product/production line quality checks. At the European level, Eurostat reports that by 2025, 19.95% of EU companies with at least 10 employees will use at least one AI technology, while technologies closest to computer vision (image recognition/processing) are adopted by between 3.78% and 7.22% of companies. In Italy, according to ISTAT 2025, 16.41% of companies use at least one AI technology, but “image recognition/processing” technology is adopted by just 2.91% of companies—a figure that highlights significant untapped potential.
Bonfiglioli Consulting’s Benchmarking Study Operations 2025 invites us to interpret these rates in terms of overall maturity: a PoC is not enough; a Lean&Digital operational model is needed—supported by data, integration, and expertise—to transform computer vision from an isolated initiative into a system-wide capability.
An effective roadmap follows this sequence and does not skip the intermediate steps:
💡 Bonfiglioli Consulting supports manufacturing companies at every stage of this journey, from initial diagnosis to industrialization. Contact us for an assessment →
Computer vision can become a pillar of Zero Defect manufacturing only when treated as an integrated three-tier system: a measurement system with calibrated sensors, technical standards, and metrological validation; a decision-making system with clear rules, defined responsibilities, and integration with the QMS; an adaptive system with continuous monitoring, controlled updates, and structured AI governance.
To move from “quality control” to "quality by design," the starting point is not the model, but the data–standard–process system
Computer Vision is a technology based on industrial cameras and artificial intelligence algorithms that enables the automatic inspection of products during production, detecting defects, dimensional anomalies, or non-conformities in real time. Unlike human visual inspection, it ensures repeatability, speed, and traceability of results.
Computer Vision eliminates the variability of human judgment, reduces scrap and complaints, cuts down on quality-related costs, and enables full traceability of every inspection. According to Google Cloud data, 39% of manufacturers already use it for quality inspection. Compared to manual inspection, it guarantees 100% coverage of produced parts, 24 hours a day.
The cost varies depending on the complexity of the inspection scene, the number of cameras, the type of defects to be detected, and the level of integration with PLCs, MES, and QMS. Simpler projects start at a few tens of thousands of euros; solutions based on Deep Learning and multi-station setups require more significant investments, but often yield an ROI of less than 12–18 months thanks to reduced scrap and rework costs.
The main regulatory and technical references are: EMVA 1288 for the characterization of industrial cameras, GenICam for the standard camera interface, OPC UA for Machine Vision for IT/OT integration, and the quality system standards ISO 9001, ISO/IEC 42001 (AI management), and ISO/IEC 23894 (AI risk management). Compliance with these standards ensures metrological validation and system governance.
Zero-Defect manufacturing is not achieved with a single algorithm, but by building a three-tier integrated system: a measurement system (calibrated sensors, technical standards, validation), a decision-making system (clear rules, integration with the QMS, defined responsibilities), and an adaptive system (drift monitoring, controlled updates of AI models, structured governance). The starting point is always the data–standard–process system, not the technology itself.
ai translated