Computer vision and artificial intelligence: innovation or consolidation of a mature technology? Cost or opportunity?
If you are considering whether to introduce machine vision systems into production, the right question is not “are we ready for this technology?”. The correct question is: “how much are we missing by not having it yet? Can I quantify how much it is costing me to forgo implementing computer vision and artificial intelligence systems?”
Why machine vision is already a mature technology
Machine vision is certainly not new. Machines have been “seeing” and controlling product quality for over fifty years. It is one of the founding technologies of industrial automation, consolidated, reliable, and present in production lines worldwide—from automotive to food, from pharmaceuticals to electronics. It is not a bet on the future: it is a tool that works today, has already proven its value, and is used daily by thousands of companies to produce better, faster, and with less waste.
What does introducing computer vision and artificial intelligence systems concretely bring to production?
In truth, a world of indispensable advantages opens up. A few examples? Robots that can pick parts without the need to build, install, and manage fixtures, continuous quality control without lapses in attention, without dedicated personnel operating on one or more shifts, without human variability. Precise measurements on every part, not by sampling. Automatic traceability. Maximum flexibility and efficiency.
In this article we wish to provide clarity, in order to understand what classical vision is, where AI comes into play, which applications are already ready for your line, and what you need to know before choosing a system, including aspects that “sellers of everything and nothing” rarely tell you.
A technology that precedes artificial intelligence
There is a widespread misconception, unfortunately even among industry professionals, that tends to conflate computer vision and artificial intelligence as if they were the same thing. In truth they are not, and the distinction is not merely academic: it has concrete implications for technological choices, implementation costs, and required competencies.
Machines could “see” well before the arrival of artificial intelligence systems suitable for industrial application. In the early 1970s, computers began using specific algorithms to process images and recognize basic features: detecting edges for component positioning, identifying color differences indicative of defects, recognizing clusters of connected pixels indicating the presence of a hole.
We must wait until the 1980s to see the first practical applications. In the 1990s the first standard frame-grabbers appeared: cards for connecting cameras to a PC and processing their images. Vision systems acquired increasing functionality and robustness, definitively abandoning the experimental aspect of the previous decade, especially in the industrial field.
The fundamental point is this: machine vision is one of the founding technologies of industrial automation. For decades it has contributed to improving machine flexibility, facilitating robot implementation, increasing product quality, and optimizing manufacturing and logistics. Only in recent years has it begun to merge with AI, expanding the potential of two technologies that mutually amplify each other: computer vision and artificial intelligence
Classical vision: deterministic algorithms without AI
Classical machine vision is based on operations that do not require any artificial intelligence component. The heart of these systems consists of deterministic and repeatable algorithms. Typical operations include:
- Edge detection — detection of edges and contours for localization and measurement
- Blob analysis — analysis of connected areas for counting, shape, and size
- Pattern matching — comparison with reference templates for presence/absence verification
- Geometric measurements — distances, diameters, angles, eccentricity
- Code reading — barcode, DataMatrix, QR, OCR
The image, captured and made comprehensible to the computer, is processed with algorithms capable of identifying relevant features such as contours, edges, shapes, structures. These are amplified in order to perform the checks for which the system was designed. Based on the results, the system produces an output: it can be a quality decision on a part, the reading of a code to communicate to the MES, the detection of dimensions that guide a robot in picking, or many other things. And all this was already possible before the advent of AI.
The advantages of classical vision compared to approaches that combine computer vision and artificial intelligence are concrete and in many contexts still decisive:
- Completely transparent and explainable behavior
- Predictable and verifiable results a priori
- No training dataset required
- Rapid commissioning cycle
- Low processing latency, compatible with high production rates
- Ease of certification in regulated environments
- Certain results and easily replicable applications
The evolution toward AI: when and why
The integration process between computer vision and artificial intelligence has not supplanted classical vision: it has complemented it to address problems that traditional algorithms could not manage satisfactorily.
Although experimentation with neural networks in machine vision predates 1996, it is precisely from that year, especially in Europe and Japan, that the first industrial applications capable of brilliantly solving complex problems appeared. However, the difficulty of training and the not always consistent quality of results long hindered widespread adoption. The models produced by neural networks, while very efficient, are not explainable in human symbolic language: the results must be accepted “as they are,” hence the definition of black box. Of particular interest is the work of Benitez, Castro, Requena.
The decisive breakthrough came in 2012, when AlexNet, a convolutional neural network, won the ImageNet competition overwhelmingly, marking the beginning of the modern era of computer vision based on deep learning. The famous paper is downloadable in PDF at this link.
From that moment, AI has made previously unthinkable applications possible. A system based on neural networks, for example, is capable of identifying and classifying surface anomalies that traditional methods could not manage with sufficient reliability. And above all, it radically changes the configuration paradigm: it is no longer necessary to program the system with the parameters of conforming and non-conforming parts. It is sufficient to provide labeled images of good parts and scrap parts: the neural network is trained and the system is operational.
Computer vision and artificial intelligence: applications in production.
The current landscape: quality control and defect inspection
This is the most widespread application featuring computer vision and artificial intelligence: no industrial production can claim to be free from defects, and most of them are visually detectable. The main quality controls fall within the typical functionalities of vision: surface inspection, verification of presence and absence of components, counting, color verification.
The choice between classical vision and AI depends on the nature of the defect:
- Classical vision → geometrically defined defects, absence of components, codifiable color variations, measurements and counts
- AI / deep learning → morphologically irregular defects, non-uniform surfaces, anomalies difficult to classify a priori
Measurement and dimensional verification
Computer vision and artificial intelligence systems allow one- or two-dimensional measurements—diameters, lengths, heights, eccentricity, linearity—up to three-dimensional measurements with volume calculations through images acquired from multiple angles. Vision is particularly suitable for measuring fragile objects or those difficult to reach with contact measurement systems.
Counting and localization
Image processing algorithms identify individual products regardless of their position and count their passage on production lines, eliminating mechanical contact systems subject to wear. These are applications entirely based on classical vision, consolidated and reliable.
Robot guidance and pick & place
A significant portion of computer vision and artificial intelligence systems is dedicated to robot-interfaced systems. We specifically refer to loading and unloading applications in which the robot must identify the exact position of objects to be picked from a conveyor or bin, dynamically adapting to their random arrangement.
Identification and traceability
Computer vision and artificial intelligence systems read 1D barcodes, DataMatrix, QR Codes, and alphanumeric text in OCR, including codes marked directly on the component (Direct Part Marking). In this area too, AI has made a concrete contribution: the neural network integrated into modern readers significantly improves reading reliability on difficult surfaces, critical angles, or partially degraded codes.
Architecture of computer vision and artificial intelligence systems: from hardware to algorithms
The heart of the computer vision and artificial intelligence system consists of one or more cameras paired with appropriately chosen light sources, to which is added the system’s intelligence: processing algorithms and management software. The design is extremely flexible, as the choice of components is closely linked to production requirements and analysis needs.
The main available hardware architectures are:
- Smart camera — compact system with integrated processing; suitable for simple tasks, limited spaces, and contained budgets
- Industrial PC + separate camera — greater processing power, ideal for complex applications or multi-camera configurations
- Embedded systems — low-consumption solutions for distributed installations on the line
The main global manufacturers offer both types with software toolboxes that combine classical vision tools and AI modules in the same development environment. The same system can therefore integrate traditional tools and AI tools, choosing the most suitable approach for each problem. After 2022 these hybrid systems have reached a level of maturity that makes them today the standard choice for most industrial applications.
Classical vision or computer vision and artificial intelligence systems: how to choose
The practical question that every technical manager must ask is not “do we use AI?” but “which approach is most robust for this specific problem?”.
At Trebi we have developed over the years solid experience in designing machine vision systems. We have long been effectively integrating computer vision and artificial intelligence in many of our projects, but not in all. Having a new technology available does not mean introducing it indiscriminately. In other words, Trebi’s approach is always consultative: we design efficient and effective solutions, we do not sell anything off-the-shelf.
For this reason we can afford to provide some advice that, at least in general terms, can be useful for better orientation, and to concretely explain our way of thinking, designing, and supporting the Client.
Choose classical vision when:
- The defect or feature to be detected is well defined and geometrically codifiable
- Product variability is low and controlled
- Full transparency and explainability of the decision-making process is required, for example in certified or regulated contexts
- Cycle times are very fast and the inference latency of a neural network would be incompatible with production rate
- A dataset of labeled images for training is not available
Choose AI when:
- Defect variability is high and difficult to describe with deterministic rules
- The product has irregular surfaces or non-uniform appearance
- You want to reduce configuration and commissioning time on new products
- The features to be extracted are complex and not manually codifiable
- Training procedures can be performed effectively
In both cases, the advantage over human inspection remains unchanged: machines do not tire, do not get distracted, and maintain the same precision on the first part as on the millionth.
What vision system sellers don’t tell you
Those who sell industrial vision systems, regardless of the combination of computer vision and artificial intelligence concepts, tend to present demos under optimal conditions: perfect lighting, stationary parts, neutral background. The reality of a production line is radically different, and the problems that emerge during implementation are rarely anticipated during commercial negotiations.
Lighting is half the system—perhaps even more. A camera worth thousands of euros paired with inadequate lighting produces worse results than an inexpensive camera with good lighting. Vibrations, thermal variations, dust accumulation on optics, reflections from shiny metal surfaces, variations in ambient light: all these factors degrade over time the performance of a system designed under laboratory conditions. In many cases a project proves unfeasible not due to camera or software limitations, but because there is no way to properly illuminate the parts.
At Trebi we have been carefully studying the lighting problem for years, which if not properly managed can cripple any attempt to combine computer vision and artificial intelligence. From our experience we can synthesize a reference framework that we believe can represent a valid starting point for beginning to understand the complexity of the problem.
Types of lighting in industrial vision systems
The choice of illuminator is a technical decision in all respects, also preparatory to the implementation of any system that combines the concepts of computer vision and artificial intelligence. The main solutions differ by geometry, spectrum, and operating mode.
By geometry:
- Diffuse front lighting — illuminates the part from multiple angles, reduces shadows and reflections; suitable for opaque surfaces and presence/absence checks
- Grazing light (darkfield) — light beam at low angle to the surface; highlights scratches, scoring, and surface defects invisible with front lighting
- Backlighting (backlight) — the part is captured against the light; ideal for precise dimensional measurements, hole detection, profile and shape verification
- Coaxial lighting — light projected along the same optical axis as the camera via beam splitter; eliminates reflections on flat and shiny surfaces such as glass, mirror-polished metal, or wafers
- Polarized light — crossed polarizing filters on source and lens suppress direct specular reflections, allowing only light that has interacted with the material to pass. Preferred solution for painted surfaces, plastics, glass, and displays; widely used in automotive and touch screen inspection
By spectrum:
- White light — general use, good color rendering
- Monochromatic light (red, green, blue, infrared) — each wavelength interacts differently with materials; red light penetrates translucent materials, blue light increases contrast on light surfaces, infrared reveals features invisible to the human eye
- UV light — excites fluorescence in some materials; useful for detecting traces of adhesives, paints, or organic contaminants
By operating mode:
- Continuous lighting — simple and economical, suitable for slow lines
- Stroboscopic lighting — flash synchronized with acquisition freezes part movement, eliminates motion blur on fast lines and extends LED life; standard solution in high-rate conveyor installations
The wrong choice of illuminator is one of the most frequent causes of systems that work perfectly during testing and fail in production. The practical rule is simple: choose the lighting first, then the camera.
Computer vision and artificial intelligence: food for thought.
Why has Trebi never adopted a “seller’s” approach, instead orienting itself toward results-oriented design, pursuing efficiency and effectiveness objectives on behalf of its Clients?
Why has Trebi not exclusively and uncritically embraced the new systems that combine computer vision and artificial intelligence, abandoning classical machine vision?
We could simply answer “because we don’t hide behind our finger,” but that would be disrespectful of our experience and the value we actually give to our work and especially to yours.
Here then are four clarifications that will certainly make you reflect, concerning aspects that are always present, addressed, and resolved in our designs.
1 – Real product variability is always greater than expected. Samples presented during the validation phase often belong to the same batch, with homogeneous characteristics. In production, parts arrive with variations in surface finish, tolerances at the limits, machining residues, handling scratches. This variability makes setup long and complex if you are not experienced in managing it.
2 – AI systems require data, time, and expertise. A model that reliably integrates computer vision and artificial intelligence requires a significant number of labeled images for each defect class, including images of rare defects that by definition are difficult to collect. The setup process can be lengthy and the company that does not have internal expertise risks becoming dependent on the supplier for any model updates.
3 – Integration with existing systems is almost always more complex than expected. A vision system is not an island: it must communicate with the line PLC, with the MES, with traceability systems, and with quality databases. Communication protocols, response times, alarm management, archiving of images of rejected parts: every aspect requires integration engineering that is rarely included in the base price.
4 – Awareness of these aspects should not discourage technology adoption. The benefits from integrating computer vision and artificial intelligence remain real and documented. They should guide toward a rigorous evaluation of solutions, a pilot phase conducted under real production conditions, and the choice of a technical partner evaluated not only on price, but on the ability to provide support over time.
Conclusions
The implementation of computer vision and artificial intelligence systems in manufacturing industry represents a mature discipline, with roots in the 1970s and practical applications consolidated since the 1980s and 1990s. All this well before artificial intelligence became accessible to companies. Today it is one of the key technologies of industrial automation, with applications across all manufacturing and process sectors.
Computer vision and artificial intelligence: key messages.
- Classical vision is still today the correct choice for a wide range of standard industrial applications
- Integrated computer vision and artificial intelligence systems have expanded the application scope, without replacing the traditional approach
- The most effective solutions combine the two approaches, choosing from time to time the most suitable for the problem
- The technological choice must start from the problem, not from the technology or the abstract belief that combining computer vision and artificial intelligence automatically brings advantages.
- Consolidated experience, design, and implementation capabilities are needed to profitably use these systems.
At Trebi we installed the first machine vision systems in 2002, convinced even then that this technology would represent a strategic direction for industrial automation. Today we install classical vision systems and computer vision and artificial intelligence systems on a daily basis, studying each application in a customized manner, to concretely improve the productivity of our Clients.
