- Stellenbosch University student Johannes le Roux devoted his thesis to teaching an AI model how to assess a steak like a braai master.
- He built a visual image-processing and machine-learning model that could estimate a steak’s internal temperature without cutting it or using a thermometer.
- His supervisor said what started as a joke turned into one of the most complex research topics his final-year engineering students have done.
- For more financial news, visit News24 Business.
Stellenbosch University student Johannes le Roux was excited to learn that he could earn his honours degree in civil engineering by braaiing steak.
Naturally, there was a little bit more to it than that.
Le Roux’s half-year thesis in his honours year looked into whether it is feasible for a non-intrusive, AI-assisted system to predict the “doneness” of a steak.
Several pages of this thesis unpack what “steak doneness” means. There is much more to it than you might think, as the internal temperature, browning, moisture, and grill marks on a steak all play a role in a consumer’s perception.
Essentially, Le Roux developed a visual image-processing and machine-learning model that could estimate a steak’s internal temperature without cutting it, poking it with a thermometer, or relying on gut feel.
He called his creation BraAI (which he pronounces braai AI).
He said he would love to take credit for the idea, but it was actually one of his professors who put it forward as a topic for the half-year thesis that all civil engineering honours students at Stellenbosch University must complete.
“It just seemed like such a good opportunity to do five months of research on something that has got nothing to do with your degree, and broaden that engineering degree even more.”
He said he also had an ulterior motive. “To be honest, I knew I was going to have to make a dataset, and I thought hopefully I could get some steaks out of it.”
How BraAI came to be
The first thing Le Roux needed was some data. He said that there were no datasets that he could find on “steak doneness”, which meant he had to go out and make his own.
The data he needed was what he calls “image, temperature pairs”.
If he could teach a model what a steak looked like at different internal temperatures, he could then use that to estimate the temperature of the steak using images.
Le Roux connected thermometers, called thermocouples, to a steak, which he cooked on a braai. The thermocouples measured the temperature inside the steak and on the grill. He also had a camera positioned above the braai, capturing pictures of the steak every 10 seconds.
He recorded about 800 image, temperature pairs while cooking steaks.
Le Roux said that the university paid for 20 sirloin steaks, but he only needed 10 to generate his dataset.
He (and his friends) made good use of his dataset. “It was a good month of eating,” he said.
He said that once he had collected the data, he performed what is known as a data split. Essentially, some image, temperature pairs are used to train models, while others are used for validation.
This allows you to teach the model what to look out for and test its performance on data it hasn’t seen.
How well did it work?
Le Roux built three models which estimated the internal temperature of the steak at different points in the cooking process:
3 models
- A metadata-only model, which made an estimation based only on how hot the grill surface was and how long the steak had been cooking.
- An image-only model that completely ignored the temperature data that had been collected. This model was trained to estimate the steak’s internal temperature using only visual cues, such as its browning. Le Roux initially wanted the model to also assess steak shrinkage, as that is a useful indicator of doneness, but that proved technically challenging.
- A hybrid model that combined data from the images, the length of time the steaks had been on the braai, and the grill temperature to predict internal temperature.
To test how well the models worked, Le Roux compared the steak’s actual internal temperature to each model’s prediction. The difference between predicted and actual temperature is called the mean absolute error.
He tested the models in two ways: predicting temperatures at unseen moments from the same steak and predicting temperatures for entirely new steaks.
For reference, there is usually a 5°C difference in the internal temperature of a medium-rare steak compared to a rare steak.
The image-only model was the best at predicting the internal temperature of steaks it had seen, with an average error of 1.81°C; however, it was the worst of the lot when predicting a steak it hadn’t been trained on, with a mean absolute error of 5.63°C.
The metadata model was quite consistent across both categories, with an average difference of just more than 4°C.
The hybrid model, which combined visual data, grill temperature and the length of time the steak had been on the braai, was quite accurate and consistent. It was off by an average of 2.56°C when looking at data from one steak, and 4.4°C when looking at steaks it hadn’t seen.
Le Roux ultimately concluded that the BraAI hybrid model was a practical path to reliable real-time doneness estimation under variable outdoor conditions, such as on a braai.
Potential
Le Roux was recognised for his work. He got 80% on his thesis, which he was pretty chuffed with, and was nominated to represent the civil engineering department at an innovation awards evening hosted by the Stellenbosch University Faculty of Engineering.
He has now landed a job at a structural engineering firm. He won’t be directly applying what he learnt about computer vision and AI in his new role, but would like to continue learning about the field.
Le Roux said for him, this wasn’t just another university project, and he became quite obsessed with it once he realised the potential.
“When you really start to think about what the model is actually doing, it is taking an object; the object goes through a process, and then there is an end product, and the whole point is that you don’t want to put something in the object while it is going through the process. You don’t want to cut it to see the steak,” said Le Roux.
He said when you start thinking about it that way, there is nothing stopping you from applying the same concept to many other industries. You could follow a similar process to predict the internal temperature of a brick being fired in a kiln, for example, to detect any issues in the process.
Richard Walls, Le Roux’s thesis supervisor and a professor of structural and fire engineering at Stellenbosch University, said the project was highly innovative.
“What started out as a research team joke turned into one of the most complex research topics my final-year B.Eng students have done. Johannes’ work combining machine learning, visual image recognition, temperature measurement and people’s cooking preferences was highly innovative.”
He said there were already ideas to take the idea and enhance it and to work with industry to “produce great braais”. “We also have an endless supply of students who can assist with quality control during cooking research.”
3989
