Quick answer
Machine learning can spot visual defects in roasted coffee beans, such as quakers, under-roasted, burnt, insect-damaged and mould-damaged beans, by training image classification models on photos of them. In Ivan Barinskiy's University of York project, which we helped with alongside Vibe With Coffee, the pre-trained MobileNet and ResNet 50 models worked best, with the highest overall accuracy of 95%. The full paper, dataset and code are all open source.
Coffee Defect Detection Collaboration
Back in 2023, Ivan Barinskiy, a Computer Science student at the University of York mentioned that he was considering implementing a machine learning model to detect roasted coffee bean defects, and reached out to me to see if I would be willing to contribute to his third year dissertation alongside Nottingham based coffee roaster, Vibe With Coffee.
After talking through the details, I started saving every coffee bean defect that I could get my hands on, focusing specifically on Quakers, Under-roasted coffee beans, Burnt coffee beans Insect Damaged coffee beans (such as coffee borer beetle damage) and Mould Damaged coffee beans - basically any defect that can be identified visually.

Coffee Beans Evaluated
Lots of coffee beans were sent for evaluation:
- Wush Wush
- Caturra
- SL-28
- Ethiopian Heirloom
- Catuai
- Red Bourbon
- Typica
- Guji & Yirgacheffe Landraces

And included many different processing methods:
- Carbonic Maceration
- Fruit Infused
- Honey
- Natural
- Anaerobic Natural
- Washed

If you are interested in how coffee variety can influence flavour and visual appearance, you may also enjoy our guide to Pink Bourbon coffee.
Machine Learning Method
Ivan's implementation method involved training several image classification algorithms to determine which one worked best.
In his own words, Ivan describes them as the following:
-
KNN-based Classifier: The images colour histograms were compared using the Canberra distance metric (KNN Stands for K-Nearest-Neighbours)
-
MobileNet V2: This model was trained from scratch specifically for this task.
-
Pre-trained MobileNet: This model had already been trained on a large set of general images (the ImageNet dataset) before being fine-tuned for coffee beans.
-
50-layer ResNet: Another model pre-trained on the ImageNet dataset, known for its depth and ability to handle complex image recognition tasks.
Initial Results
Here are the initial results from Ivan's trial:

Which demonstrated that the Pre-trained MobileNet and ResNet 50 were the most effective at detecting defects, with the highest overall accuracy of 95%. After some training renditions, Ivan's machine learning models MobileNet and ResNet 50 models enabled him to produce an incredibly accurate set of results.

Full Publication and Dataset
The Full publication of Ivan's work can be found here:
Click here for the full publication
Along with the full data set, which can be found here:
Barinskiy, I. (2024). Identifying defects in roasted coffee beans using image classification algorithms [Data set]. https://github.com/dont-text-me/RoastedCoffeeDefectDataset
You can find the abridged data here:
https://bean-classifier-demo.vercel.app/
And finally, you can find the full open source publication including the relevant download links here: https://github.com/dont-text-me/PRBX-work
Don't forget to check out our current coffee offerings to get your hands on some of the best beans out there (Defect free ;) )
Coffee Defect Detection FAQs
Which coffee bean defects can machine learning detect?
Basically any defect that can be identified visually. The project focused on quakers, under-roasted beans, burnt beans, insect-damaged beans (such as coffee borer beetle damage) and mould-damaged beans.
Which model was best at detecting coffee defects?
Ivan trained several image classification models: a KNN-based classifier comparing colour histograms, MobileNet V2 trained from scratch, a pre-trained MobileNet and a 50-layer ResNet. The pre-trained MobileNet and ResNet 50 were the most effective, with the highest overall accuracy of 95%.
What coffees were used to train the models?
Lots of different beans: Wush Wush, Caturra, SL-28, Ethiopian Heirloom, Catuai, Red Bourbon, Typica and Guji and Yirgacheffe landraces, across carbonic maceration, fruit infused, honey, natural, anaerobic natural and washed processing.
Can I download the coffee defect dataset?
Yes. The whole project is open source. You can read the full publication, browse the abridged data, and find the full project with download links on GitHub.
Who carried out the research?
Ivan Barinskiy, a Computer Science student at the University of York, as part of his third year dissertation. We contributed to the project alongside Nottingham based coffee roaster Vibe With Coffee, saving every visually identifiable defect we could get our hands on.
Work With Us
Looking to collaborate with us on a project like this one?
Drop us an enquiry here!












