Cool Pear Selection: How AI Prevents Food Waste

Less Manual Labor, Less Food Waste, More Quality. Former data science student Cees Maessen explored in his master’s thesis how artificial intelligence can be used to sort pears more efficiently: ‘We don’t want bad pears on supermarket shelves, but we also don’t want good pears ending up in the trash.

Cees Maessen: ‘The idea of full traceability, from tree to pear, is still a distant prospect, but not unimaginable.’ Image: Jack Tummers

What is your thesis about?

‘My thesis focuses on making the pear sorting process more efficient and consistent. When pears are harvested, not all of them are suitable for sale in supermarkets. Some have defects such as dents, black streaks, or diseases. The sorting process is carried out by specialized companies, not by the pear growers themselves.

‘These companies use conveyor belts and water channels to transport pears to sorting machines. While machines play a role, manual labor remains essential. Workers visually inspect the pears, which is time-consuming and prone to errors, especially given the variability in worker skills. Some employees have been doing this work for a long time and are highly skilled, but temporary workers hired during harvest season may sometimes struggle to accurately judge which pears are suitable or not.

‘The challenge is to exclude bad pears while avoiding the rejection of good ones, as this leads to food waste. We don’t want bad pears in supermarkets, but we also don’t want good pears ending up in the trash. How can computer vision based on artificial intelligence (AI) be used to improve this process? That’s what I wanted to find out.’

How did you approach your research?

‘I started by creating an infrastructure for collecting and processing data. The sorting process uses advanced cameras that capture pears from different angles. The challenge was to analyze these images with minimal manual input. To achieve this, I used an AI technique that learns from unlabeled data. This means the models independently recognize patterns without requiring all images to be manually labeled beforehand.

‘In the first phase, the model learned the general characteristics of pears, such as shape, color, and texture. Abnormal objects, such as severely misshapen pears or defects, were automatically separated from normal pears. Then, I trained the model with labeled data to recognize specific categories, such as ‘rotten’ or ‘not rotten.”

Cees Maessen: ‘I trained the model with labeled data to recognize specific categories, such as ‘rotten’ or ‘not rotten.” Image: Cees Maessen

What were the findings of your research?

‘One key result is that AI has the potential to significantly improve the sorting process. The model I developed can more accurately determine whether a pear is suitable for supermarkets, the industry, or should be rejected altogether.

‘I also discovered that AI without labels can not only detect rotten spots effectively but also identify irregular shapes and colors that indicate diseases or damage. While AI detecting such issues with labeled data has been known for over a decade, achieving this with unlabeled data is new and exciting.

‘Another important insight is that the model handles challenges well. Unlike traditional computer vision systems, which struggle to distinguish, for example, a brown stem from a brown rotten spot, more advanced AI models can detect subtle differences. This leads to more reliable and consistent results. Additionally, the system offers potential benefits for precision agriculture. By feeding data about pears back to growers, they can better understand where problems arise.’

Why is this research important?

‘The project addresses several societal and economic issues. First, there is a labor shortage for manually sorting fruit. The work is demanding, seasonal, and becoming increasingly unattractive. Automation offers a solution here.

‘Second, it contributes to the fight against food waste. The current process sometimes leads to unnecessary rejection of good pears simply because manual inspection and traditional computer vision are not perfect. By using AI, we can make better decisions and preserve more fruit. If these techniques are further developed, they could also be applied to other crops, such as apples, tomatoes, and cucumbers. This paves the way for more efficient agricultural practices and higher-quality products for consumers.’

What does this mean for the future?

‘Although important steps have been taken, much work remains to be done. The techniques I developed provide a solid foundation for further innovations. The idea of full traceability, from tree to pear, is still a vision for the future, but it’s not unimaginable. Imagine if growers or consumers could know exactly where their fruit comes from and the journey it has taken. This would not only provide transparency but also help make agriculture even more sustainable and efficient.’

Master thesis

A literature review, lab experiments, or working with SPSS? Students at Tilburg University write the most diverse theses. In the monthly column Master’s Thesis, Univers highlights one of them.

Author: Cees Maessen
Title: Label-Efficient Pear Characteristic Classification: Evaluating Self-Supervised Learning Techniques in an Operational Setting
Supervisor: Rogier Brussee
Grade: 8.5
Master’s Program: Data Science in Business & Entrepreneurship

Advertentie.

Bekijk meer recent nieuws

Schrijf je in voor onze nieuwsbrief

Blijf op de hoogte. Meld je aan voor de nieuwsbrief van Univers.