Preventing foreign object contamination is a growing priority for food processors. According to USDA, it accounted for more than 75% of the total volume of food recalled by the Food Safety and Inspection Service in 2019. Contamination isn’t a novel issue, however, and many processors are looking for new and innovative solutions to help automate detection and increase the likelihood of finding foreign materials. One reason for this trend is that materials such as plastics and rubber are showing up with greater frequency, and these materials are often missed by metal detectors and X-rays.
Improved detection of foreign contaminants will help reduce food waste as well as lower costs and the risk of recalls. Many processors look to identify contaminants early so they can address an issue quickly and minimize the impact on production.
The good news is that detection technology is evolving quickly. Vision-based systems are a good example of growing innovation in the processing sector. But what exactly do we mean when we talk about vision systems for food processing?
While X-rays and metal detectors are commonplace in processing, vision systems are relatively new. “Vision system” is an umbrella term for a number of different systems with widely varying capabilities and characteristics. In this article, we will compare different vision systems in terms of how they function, their specific attributes, and how they may benefit food processors.
The Science of Seeing
To understand the differences among types of vision systems, it’s useful to remember how light works—the science behind how we see things.
Our eyes are only able to see three color bands: red, green, and blue, otherwise known as the visible spectrum. However, light is actually made up of thousands of different wavelengths. Each wavelength behaves differently and interacts differently with various materials. We can use these diverse wavelengths of light, both inside and outside of the visible spectrum, to gather information about different materials or objects.
When it comes to comparing vision systems, there are three main differences to consider:
- The number of light bands, i.e., the number of colors that a system is able to see (otherwise known as wavelengths);
- The spectral resolution—the higher the resolution, the smaller the ‘gaps’ between each color or wavelength; and
- The amount of information a vision system is able to see per pixel of an image. A pixel is the smallest unit of information that makes up a picture.
Together, these three characteristics define the level of detail a vision system is able to consider, the ability of that system to detect a variety of different materials, and how “trainable” a system is, i.e., whether it can learn from the information it’s gathering.
The Art of Looking
The original vision systems are our eyes. Human inspectors are frequently brought in or added when there has been a contamination event. Studies from other industries have shown, however, that after just 15 minutes on an inspection task, human performance drops dramatically. After 30 minutes on a task, the probability of detection falls by more than 50% on average, meaning that inspectors have a one in two chance of missing the materials they’ve been hired to find.
This can be due to multiple factors, including line speed, levels of training or experience, fatigue or illness, and even external factors such as background noise or lighting conditions. Studies in other industries have shown that simply adding more inspectors does not necessarily increase detection rates.
Automation of repetitive tasks—such as inspection—delivers better and more consistent outcomes. It also frees up valuable staff for more important and, often, safer tasks that require human expertise.
Camera-Based Inspection Systems
Camera-based systems are the most well-understood type of vision system. Cameras have been around for more than a century, and most of us carry one in our pocket at all times. Camera-based inspection systems are the closest in performance to the human eye, which means that they will only see objects within the three colors of the visible spectrum. Their advantage over human inspectors can be greater consistency; they don’t get tired or lose concentration. However, cameras are not effective in detecting contaminants when there is little contrast between the object being inspected and the material they are looking for—for example, white plastic on a fatty piece of chicken or on ground pork trim.
When it comes to detecting contaminants, cameras will likely miss items such as clear plastics or any objects similar in color to the product. Line speed and lighting conditions can also affect camera performance, because cameras have trouble seeing things on a messy or variable background, such as meat on a line. Figure 1 shows how a camera can more easily see objects when the background is plain.
Camera-based systems are ideal for assessing size and shape, such as with nuggets or patties.
Beyond the Visible Spectrum
Multi-spectral systems are different from camera-based systems. Instead of being limited to three colors, as in a camera-based system, multispectral systems are able to see between three and 15 spectral bands, and can see colors outside the visible spectrum. This enables them to see some chemical properties of the inspected object.
Multi-spectral systems were used in early space-based imaging to map landscape details on Earth. Detection in these systems is based on the materials the system expects to see. In the case of space-based imaging, the systems were set to detect water versus land versus vegetation. In food processing, these systems can be useful when contaminants are consistently made of the same materials; however, new or previously unknown contaminants will be missed, even if this “new” contaminant reappears multiple times.
Because these types of systems use a set number of spectral bands, they have a limited capacity to learn from what they see over time. And, like camera-based systems, multispectral systems aren’t able to assess quality measures.
From Multispectral to Hyperspectral
As the name suggests, hyperspectral systems collect information across the electromagnetic spectrum. They measure continuous bands through both the visible and invisible spectra, which means they see hundreds or thousands of essentially continuous light bands. This means that hyperspectral systems gather very robust data about the materials being inspected, down to a chemical level.
Hyperspectral imaging systems produce incredibly rich data on every piece of product they inspect. In a food processing plant, that means you can not only find, but identify, foreign materials based on their chemical signature, reducing your time to resolve issues by pointing the way to the likely source of the contaminant.
Hyperspectral imaging systems can go beyond just finding foreign materials. Unlike multispectral or camera-based systems, hyperspectral systems can assess quality measures such as steak tenderness or fat/lean ratios in sausages and can find myopathies like woody breast or spaghetti breast in poultry.
Hyperspectral systems are also exceptional in another way: These systems can use artificial intelligence (AI) to learn from the chemical data they collect over time. This makes these systems highly effective at identifying new or unexpected contaminants. It also means these systems can grow and change over time as the needs of a processing plant change, without the need for new capital equipment.
Recent advances in computing and computer processing have made it possible for these hyperspectral systems to operate on the line in real time.
How to Choose a Vision-Based System
Vision systems have tremendous advantages for food processing, but it’s important to know which system is the right one for your plant. Asking the right questions will help guide your selection process.
First, ask to see a detection curve for the system. A detection curve, a chart that shows object size plotted against probability of detection, will give you a very clear indication of how successful a system will be in detecting objects of any size. Figure 2 shows examples of detection curves for different materials identified by a hyperspectral imaging system.
A detection curve provides much more useful insight than simply asking about the smallest size of object a system can detect. A system that claims to find microscopic objects, for example, might only find them in very rare instances.
Second, ask about false positive rates. Using the same example, a system might claim to find a very high number of tiny objects. But what if many of these detections are false positives, meaning that there is no contaminant actually present? A lot of valuable product may be unnecessarily discarded.
Finally, ask if the system is future-proof. Will it be able to expand to detect new types of contaminants over time? Might you need to evaluate quality metrics in the future? Plants are constantly evolving, and new processing techniques or types of products bring in new forms of contaminants and evolving quality issues. Will the system be able to adapt to these changes?
The food processing sector is embracing innovation at a faster pace than ever before. But, as with any evolving technology, the key is understanding the differences among available detection systems. The right approach for your business may even be a combination of different systems—a multi-hurdled approach.
Asking the right questions will guide you in your selection and help drive efficiency and safety in the plant, while reducing food waste and costs in the long term.