Visual Process Monitoring in Production: What a Camera Image Still Does Not Tell You

What a camera in production can do – and what it cannot

The word camera covers technically very different systems. A conventional RGB camera captures visible light. 3D systems determine spatial distances. Thermal cameras capture thermal radiation. Multispectral and hyperspectral systems open up additional spectral ranges.

These are not merely variants of the same technology. They determine which information can be obtained from an acquisition at all.

What a camera in production can contribute therefore depends first on the information needed – not on the camera model.

In food processing, machine vision is used, for example, for automated non-destructive inspection. A recent survey describes RGB and multispectral cameras as well as depth and other sensing systems. Illumination is part of the acquisition setup too: intensity and uniformity influence whether shadows, reflections or saturation already alter the image basis. ^[1]

Hyperspectral imaging combines spatial image information with spectral information. For a given image location, this means that not only a color value but a spectral signature across many wavelengths becomes available. ^[2]

With hyperspectral acquisitions, further steps lie between image and statement. The detector initially measures spectral intensity rather than reflectance directly. The images are therefore calibrated against black and white references. For interpretation, the method then also requires a calibrated and validated evaluation model. ^[2]

The decisive point is this: better image quality does not replace the wrong sensing principle. If a relevant property is not captured in the selected measurement range, more elaborate downstream analysis cannot create it from the image.

What resolution decides

Even when the right sensing principle has been selected, the question remains how accurately a property can be measured from it.

A pharmaceutical pilot system for continuous mini-tablet manufacturing provides one example. An online camera observes the droplet during production and determines, among other parameters, size, position, solidification time and settling velocity. For determining drug loading, the camera value alone is not enough: camera-based droplet size is combined with formulation concentration measured by a UV probe. Only the combination of both measurements creates the additional quality information. ^[3]

The authors then evaluated the accuracy of the size measurement. The mini-tablet radius measured by the camera differed from the reference measurement by an average of 2.3 percent and a maximum of 4 percent. For the volume calculated from that radius, the mean error increased to 6.9 percent and the maximum to 12 percent. ^[3]

The important point is not only the figures themselves. The error triples in the volume calculation because of the cubic relationship between radius and volume. The assumption of an ideal spherical shape also influences the derived quantity. ^[3]

This turns a camera issue into a more general process issue: even a correct measurement can carry amplified uncertainty through downstream calculations.

These figures are not a general accuracy specification for process cameras. They apply to this particular experimental setup. They do, however, show why an automated image measurement does not lose its uncertainty merely because it is automated.

What perspective decides

An object has more than the side facing the camera.

For automated food inspection, a survey describes multi-view acquisition for approximately spherical products such as apples, oranges or tomatoes so that the surface can be captured from several directions. For steaks and cuts of meat, chicken or fish, it describes analysis of one or two sides. Different objects therefore require different acquisition strategies. ^[1]

How large the effect can be is illustrated by a study on yield estimation in apple trees referenced in the same survey. Imaging from two opposite sides was compared with regular single-side imaging. The reported crop-estimation accuracies were 82 versus 58 percent. This is not a statement about industrial quality inspection, but it is a concrete example that the number of viewpoints can materially change the result. ^[1]

Another setup described in the survey solves the perspective problem not by adding cameras, but by moving the object. Oranges are transported and rotated simultaneously on a roller conveyor. A fixed camera therefore receives multiple views of the same product, which are subsequently combined into one image representation. ^[1]

Perspective thus becomes part of process design itself.

The same applies to illumination. In another referenced system, an enclosure and uniform LED lighting are used to make acquisition independent of changing natural light – including operation during daytime and nighttime. ^[1]

A camera should therefore not be considered in isolation. Position, optics, illumination, working distance, object movement and visible field of view together form the acquisition setup.

What the image does not know about its environment

An image can show changes in a plant. It does not automatically know the temperature or humidity at the time of capture.

This becomes clear in a study on automated monitoring of industrial hemp in a greenhouse. The setup combines fixed cameras with temperature and humidity sensors and automates image capture. Five plants were observed for the image analysis. ^[4]

For industrial processes, the interesting point here is not a single classification percentage. With such a small experimental setup, a percentage without the sample context would suggest more generality than the study supports.

What matters is the information architecture:

The camera describes a visual condition. Other sensors describe the conditions under which that state emerged.

Only their combination expands the possible statement.

A discoloration, an unusual surface or a changed plant structure may be visible. Whether the cause was temperature, humidity, material, processing or something else is not contained in the image by itself.

More measurements of the same variable do not make a new variable visible

Multiple sensors in one process serve different purposes.

In a continuous pharmaceutical tableting process, for example, load measurements, NIR sensors, microwave sensing, a camera, X-ray sensing and further test methods were combined in one sensor network.

The authors make an important distinction: multiple measurements of the same process variable can improve the reliability of that measured variable. But they do not automatically improve the observability of another variable that has not been measured. Sensors using the same measurement principle can miss process features that only become observable through a different technology. ^[5]

Redundancy and diversity therefore answer two different questions:

How reliably do I know a measured value?

and:

Which properties of the process can I observe at all?

The camera is therefore not automatically the monitoring system. It can be one information source within a larger view of the process.

How different device classes are implemented technically

The differences between sensing principles are also reflected in currently available devices. The following selection is intended only to make different technical classes tangible. It is not a ranking or purchase recommendation.

Technical values and interfaces according to manufacturer specifications, status: October 2026. Assignment to device classes and the location of processing are our classification. The selection illustrates different device classes; it is not a purchase or performance recommendation.

The same principle applies as in the acquisition table above: the right camera does not emerge from a model ranking. It follows from the process property that needs to be observed or measured.

What only becomes useful through assignment

Even the technically correct acquisition leaves another question unanswered:

What does it belong to?

A camera image can show a state precisely and still be difficult to place later if its process context is missing.

Vollmer and Palm, both with Schuler Pressen GmbH, describe a monitoring system for a hot-stamping line in which data from the press, furnace, robotics, thermographic cameras, video cameras and other sensors are captured together. All data sources are stored with a synchronized timestamp. ^[7]

For quality-relevant physical process values, the authors go one step further: the values are not only related to time ranges but linked to individual produced parts through a unique Part ID. ^[7]

For video, the same work separately describes synchronization of video and sensor data via equal timestamps. Manual image analysis and machine-vision evaluation are also presented as two different paths. ^[7]

These findings must not be merged into a statement the source does not make: the paper does not state that the video file itself is linked to the Part ID.

A design requirement can instead be derived from the two findings:

If visual information is to be assigned to a concrete operation later, image or video data need to be embedded in the same process context with time and object references.

The relevance of individual assignment is also visible in the pharmaceutical mini-tablet study. In continuous mode, individual mini-tablets could be identified, their critical quality attributes determined and the units tracked through downstream processing steps. ^[3]

The study does not say, however, that the underlying camera image remains stored together with that individual identity.

This is the difference between visual capture and process-related information.

How object, time and relationship references are represented in a digital process is described in the article on the context model. The article on process events addresses the corresponding event structure.

Storing data is not yet a retention strategy

Storage also requires two levels to be kept apart.

Vollmer and Palm first describe a fast ring-buffer database in which complete historical production data remain available without loss of resolution for short- and mid-term analysis. For the investigated system, they report a period of about two months. ^[7]

They then describe preprocessing and the long-term storage of preprocessed data for quality assurance. ^[7]

This produces two storage levels with a processing step in between:

full resolution for short- and mid-term analysis → processing → preprocessed data for long-term storage.

The study does not answer which information actually has to be retained for how long. It provides no specific regulatory retention period and makes no statement about audit trails, electronic signatures or immutability.

A process design nevertheless still has to answer an important question:

Must the complete image sequence remain available later – or the information derived from it and assigned in a traceable way?

The answer depends on the process and the requirements for later traceability. The distinction between retaining a history and assessing its integrity is discussed separately in Audit Trail and Data Integrity.

Who evaluates the image?

Acquisition and decision are not the same thing.

Visual information may be reviewed by a person. It may be evaluated algorithmically. Or an automated method may flag a case that is subsequently decided by a person.

A pharmaceutical review distinguishes between static and dynamic image analysis. It also describes image analysis as a possible tool for PAT-based in-process control; processes such as crystallization, milling, emulsification, fluidization, granulation, tableting or coating can be followed using images, and disturbances can be detected in real time. ^[6]

The authors also place manual analysis and machine-vision evaluation side by side. ^[7]

This matters for system architecture: a camera does not determine who makes the judgment. It initially provides information.

How human review remains deliberately part of a digital process is covered in Human in the Loop.

When someone views the process remotely

The same distinction applies to remote support.

A camera image can allow a remote person to view the same process as the person on site. This creates another use case for visual process information – but not automatically an automated inspection and not automatically evidence.

The system design therefore has to answer different questions: Which view does the remote person receive? Which operation does it belong to? What role does that person have? Is the interaction only communication, or is a decision being made? And which information needs to remain available afterwards?

Remote support is therefore one possible role of visual information in the process, not the umbrella term for all camera use.

The distinction from industrial smart glasses remains clear: with smart glasses, the perspective follows the person. With a stationary process camera, the perspective stays with the process.

Where a first pilot starts

A useful pilot does not begin with as many cameras as possible. It begins with one concrete question.

Should the system check whether an object is present? Should a visible surface be assessed? Is the issue dimensional deviation, temperature distribution or a change over time? Does a person need to decide, or should a method classify the case automatically first?

Only then can the sensing principle, required perspective, useful resolution and additional process data be determined.

For an initial use case, at least four things should therefore be clearly named: the observed object, the relevant characteristic, the time of evaluation and the reaction to a deviation.

This makes it possible to test whether the camera system actually provides the information the process needs.

A camera can make a great deal visible.

Whether it becomes process information is decided outside the image.

Frequently asked questions

What is visual process monitoring?

Visual process monitoring uses image or video data to make states and changes during a process visible or measurable. What can be concluded from those data depends on the sensing principle, the acquisition setup and the assignment to additional process information.

What is a process camera?

A process camera can be understood as a camera used to observe a technical or operational process. The term does not define a specific sensing principle: depending on the task, a conventional 2D camera, a 3D system, a thermal camera or a spectral method may be used.

Which camera is suitable for process monitoring?

This cannot be decided on resolution or frame rate alone. The property that needs to be observed or measured has to be defined first. Only then do sensing principle, perspective, illumination, required accuracy and possible additional sensors follow.

Is machine vision the same as visual process monitoring?

No. Machine vision refers to machine-based acquisition and evaluation of image data. Visual process monitoring is broader: an image can also be assessed by a person or used together with other process data. The decisive question is how the visual information is integrated into the process.

Is a camera image sufficient as process evidence?

An image initially shows only the captured scene. Whether it can later be used as relevant process information depends, among other things, on its assignment to time, object and operation and on the requirements of the respective process. This article does not derive a general regulatory evidentiary effect from that.