Skip to main content
Quality inspection in vegetable processing has advanced significantly over the past decade. Camera and laser-based sorting systems can now automatically inspect and sort large product streams with a high degree of accuracy. Some applications remain challenging, however, particularly when foreign material closely resembles the product itself. To address these complex sorting challenges, processors are increasingly turning to Artificial Intelligence and Deep Learning.
Building on these technologies, Optimum Sorting introduced A-EYE, an advanced Deep Learning-powered sorting model. Designed for applications such as leafy spinach and green bean processing, A-EYE can detect defects that traditional sorting systems often struggle to identify. Want to see how A-EYE takes sorting performance beyond the limits of conventional technology? Discover the real-world applications later in this article.
Key insights:
How Deep Learning expands sorting capabilities
Traditional sorting technology relies on predefined product characteristics set by operators and software developers. These characteristics may include colour, shape or size. While this method is highly effective for many applications, it reaches its limits when foreign material closely resembles the product itself. In such cases, how can a camera tell the difference between acceptable product and unwanted material?
Optimum Sorting’s A-EYE model approaches this challenge differently. Using Deep Learning, a specialized branch of Artificial Intelligence, the system learns from large amounts of training data and real-world examples. Its decision-making process closely mirrors human inspection. An experienced quality inspector does not judge a product based on a single attribute, but on the overall combination of shapes, textures and visual patterns. A-EYE applies the same logic, analyzing all of these characteristics together to form a complete understanding of what it sees.
Built on Deep Learning, A-EYE is ideally suited to complex sorting applications, including vegetable processing. The model can work alongside existing sorting software, complementing conventional detection methods such as colour defect inspection. This allows processors to combine the strengths of both approaches within a single sorting solution. Later in this article, we take a closer look at how A-EYE is applied in leafy spinach and green bean processing.
Case Study 1: LEAF SPINACH
Weed detection
One of the key challenges in leaf spinach processing is detecting unwanted material that closely resembles the product itself. Grasses and certain weeds, such as thistles and nettles, often have the same colour and texture as spinach and can even display similar leaf structures.
This is exactly where the strength of Deep Learning comes into play. While traditional machine vision systems often struggle to consistently recognize such subtle differences, a trained AI model can learn which characteristics are typical of spinach and which indicate unwanted material. As a result, defects and contaminants that previously went undetected can now be identified automatically.
Case Study 2: GREEN BEANS
Stem-end detection
A-EYE also offers clear benefits in green bean processing. One practical application is the detection of stem ends. For conventional sorting systems, stem ends and tip ends often appear very similar, making them difficult to distinguish reliably.
For conventional sorting systems, distinguishing between the stem end and tip end can be extremely challenging. However, beans with a stem end are considered undesirable in the final product. Trained on thousands of real-world examples, the model learns to recognize these subtle visual differences, enabling much more accurate identification and removal of defective beans.
Plant stalk detection
In addition, A-EYE helps detect stalks, twigs and other unwanted plant material that may enter the product stream during your green bean harvesting.
Because this plant material often shares similar visual characteristics with the green beans themselves, including colour and shape, conventional detection methods can struggle to distinguish between the two. Deep Learning takes a different approach by analyzing the complete object rather than individual features in isolation. This enables far more reliable classification, even when plant material partially overlaps with the beans or is presented at different angles on the conveyor belt.
Purslane detection
A third and particularly interesting application is the detection of purslane (Portulaca oleracea). As regulations surrounding crop protection become more stringent and the use of herbicides continues to decline, this plant is gaining a foothold in certain green bean fields. After harvesting and processing, purslane can closely resemble green beans, making reliable differentiation a significant challenge.
Thanks to Deep Learning, the sorting system can reliably learn to distinguish these subtle differences. In the past, the presence of purslane could lead to entire batches being rejected. Today, vegetable processors can sort much more precisely, reducing product loss, improving product recovery and ensuring that a greater share of the harvested crop remains suitable for consumption.
"Deep Learning increased stem-end detection rates in green beans from 70% to 97%"
Faster model training
Before a Deep Learning model can identify defects, it must be trained using large amounts of real-world data. Thanks to modern cloud infrastructure and advanced training techniques, this process is much faster today than it was only a few years ago. New AI models can also build on knowledge from existing applications, allowing new sorting solutions to be developed in significantly less time.
The difference compared to just a few years ago is significant. Whereas developing a new model once often required months of work, a comparable application can now be up and running within just a few weeks. For example, a model developed for a new bean variety can build on knowledge previously gained from green bean detection. Existing models can also be further optimized with new real-world data, often within a matter of hours. The training process typically consists of three steps:
Collecting and labeling images
- Thousands of images of both good product and defects are collected and labelled. This allows the model to learn which objects belong in the final product and which should be removed.
Training and validating
- Using these examples, the model learns to recognize patterns and distinguish between different types of objects. It is then tested using new images it has not seen before to verify its accuracy and reliability.
Real-time sorting in production
- Once validated, the model is integrated into the sorting machine. During production, it analyzes each product within a fraction of a second and automatically determines whether it should be accepted or rejected.
More quality, less waste and higher capacity
With A-EYE, Optimum Sorting significantly expands the capabilities of modern sorting technology. By combining AI and Deep Learning with camera-based inspection, and in the future with laser technology and other sensors, the system can detect defects that were previously difficult or even impossible to identify automatically.
This opens the door to applications that have traditionally been challenging for conventional machine vision systems, such as distinguishing weeds from spinach, identifying purslane among green beans, or detecting stem ends, stalks and other unwanted plant material. As a result, processors gain greater control over product quality, recover more value from their harvests and reduce product loss.
The benefits go beyond improved defect detection. Trials conducted at an Optimum Sorting customer site have shown that A-EYE can also increase sorting capacity. At one green bean processing plant, A-EYE increased throughput by +25%, raising capacity from approximately 8 to 10 tonnes per hour. The system continued to accurately identify and assess products, even when they partially overlapped on the conveyor belt. While conventional algorithms often lose accuracy under these conditions, A-EYE maintains a consistently high level of sorting accuracy. More throughput without compromising quality. Isn’t that what every processor is looking for?
”We don't see Deep Learning as an end in itself. We see it as a practical way to solve sorting challenges that were previously difficult to automate. Since 2019, we have continuously expanded our expertise in this technology. Over the past year, we have worked closely with specialised partners to advance Deep Learning applications for the fresh produce and vegetable processing industries. Today, we have the knowledge, tools and experience to tackle sorting challenges that would have been out of reach only a few years ago. With A-EYE, we aim to set a new benchmark for what processors can expect from modern sorting technology.
Jan De JongheCTO Optimum Sorting
Interested in a free demo?
Get in touch with the sales team of Optimum Sorting and have your products tested at one of our demo centers around the globe.
Experience firsthand the benefits our machines can bring to your business!
What is A-EYE?
With A-EYE, Optimum Sorting introduces a new approach to automated sorting. Powered by Deep Learning, the technology can identify complex defects that are often challenging for conventional sorting systems to detect. By learning from real-world examples rather than relying solely on predefined rules, A-EYE can recognize subtle differences that previously went unnoticed. The result is more reliable sorting, less product loss and higher product quality.
