AI for post-harvest sorting and packing: understanding the options for your business

SmartSort: Dive deeper into machine learning

SmartSort machine learning uses algorithms that detect patterns in data from past examples. For food growers, packers, and processers, machine learning is ideal for:

• Identifying foreign materials
• Detecting structural product defects (non-size related)
• Spotting color-related issues (like black spots on green)

When we talk about machine learning at TOMRA Food, we're talking about a wide spectrum of both models and applications that range from simple predictions to multivariate decision making. In other words, very simple programs all the way through to a program that takes many different considerations into account when it makes a prediction. 

In summary, it can be incredibly precise, which empowers your expert operators. When you interact with machine learning within the TOMRA 5B or 5C, you're able to set the tolerances on shape, size, and biological characteristics that set your business apart and make your customers happy -- and you can trust that the machine will do all the work for you. As you run the machine, the model learns on site and adapts with your operation. In time, it enhances the performance of your operators and can even open up new revenue streams. 

Customers who use machine learning built into the 5B and 5C find they're able to easily exceed even the strictest product safety standards and quickly adapt to meet changing customer requirements and expectations. 

Machine learning is also built into the 3A and 5A, and it's the processing power of these machines. At Sackett Ranch, the team uses a combination of the TOMRA 3A and TOMRA 5A to be able to tell their customers the size and profile of every single load. Their 3A is ideal for eliminating foreign materials, like stones, corn cobs, dirt clods, and green potatoes before storage. Then, the 5A serves as a final inspection tool: the multivariate classifier (a complex form of machine learning) then categorizes potatoes by size and quality and eliminates green and undesirable potatoes to meet the strict 2% threshold of green, undersized, oversized, and potatoes with defects.

What is deep learning?

Deep learning represents the most advanced form of machine learning currently available on the market for food growers, packers, and processors. It uses more complex neural networks to deliver superior consistency, simplicity for operators, and high-performance grading. At TOMRA Food, we use convolutional neural networks (or CNNs) to build our flagship deep learning product called LUCAi™. CNNs use layered filters to “scan” images and extract meaningful features—textures, edges, colors—automatically.

To understand convolutional neural networks, it's simplest to think of them as a structural network that is inspired (but doesn't replicate) by how the human brain processes visual input.

If you have unimpaired vision, consider your own experience of standing in a field. When you look around, you don't see soil, plants, and weather in isolation. Your eyes see these things, and your brain sees that the soil is dry, the weather is sunny, and the plants are wilting. So, you can predict that you'll need to water your field. CNNs allow our machines to think in a similar way: instead of seeing dry soil, sunny weather, and wilting plants individually and largely still reaching the decision that you need to irrigate based on the wilting plant alone, the deep layers and training data that comprise a CNN will better understand why the plant is wilting and have not only greater confidence in the prediction but also understand the severity of the wilting leaves.

Applying deep learning to predict the severity of a defect or a blemish is unique to TOMRA, and it unlocks even greater precision and customization for growers and packers.

Why a convolutional neural network compared to an alternative? They just so happen to be perfect for food sorting and grading because the nature of these complex models offer:

• Precision: Perfect for identifying blemishes, bruises, and subtle defects
• Layered Learning: Understands both simple and complex visual patterns
• Adaptability: Handles different varieties, lighting conditions, and growing environments
• Reliability: Delivers consistent results—even across shifts and operators

Traditional machine learning vs. deep learning: What’s right for your operation?

Both machine learning and deep learning offer real precision and ease-of-use for your sorting and grading operation. They're equally suited to detecting foreign materials, environmental materials and defects and helping you get the right product in the right pack.

But how do you compare?

Both are incredibly powerful tools, but they serve different needs. Machine learning is extremely powerful, flexible, and operator-friendly. Deep learning can add even more precision, but it also brings consistency, and reduced training needs both on the technology and overall.

Are you a TOMRA customer? You're likely already using AI. Stay up-to-date with the latest technology advancements in TOMRA Food. 

​TOMRA Food is committed to expanding the reach and capabilities of its deep learning models. New technologies are emerging faster than ever, and our focus remains on helping customers like you improve precision, enhance usability, and stay ahead of change. Reach out to your account manager or click the Contact Us button to send us a message.

Learn more about TOMRA's AI-enabled sorting and grading machines