Using AI to brew a perfect cup of coffee

There are as many different potential flaws in a bean as there are delicious flavor profiles in a finished, quality-controlled cup of coffee. A given bean might be withered, chipped, or greenish and unripe. It might be perforated by insect bites, marred by fungus, or contaminated by foreign matter.  

All these defects and more must be carefully screened out by the manufacturer, and existing products that have been built to aid in the process can only make comparatively rudimentary discriminations based on color and shape.  

Until recently, this has meant an exhausting manual process with imperfect results. That’s a problem, especially since in the case of some types of defects, a few troubled beans can foul the taste of a whole batch. 

Now, though, researchers at Yuan Ze University in Taiwan and collaborators in Peru have developed a way to bring the power of augmented intelligence to the rescue, with advancements powered by Lenovo’s high-performance computing system and scaffolded by support from Lenovo’s AI Innovation Center. 

“It’s very difficult and tedious work,” said Professor Yun-Chia Liang, Yuan Ze University. “We are not trying to replace workers but trying to develop a system to help them.” 

Peruvian collaborators Prof. César Beltrán Castañón and Dr. Christian E. Portugal Zambrano 

helped lend insight into the arduous bean-sorting process. 

“Every time they have the harvest for the coffee beans, they have an intensive labor over there. Everyone just sits in a row and tries to pick the defective ones manually. So, we were thinking, if we have a chance to develop a system like this, we definitely can help them to enhance the quality of inspection,” Liang explained. 

Though complex to develop, the system is relatively simple to understand. Beans go into a feeder where their images are captured by a camera. Next, deep learning AI is used to label them, quickly determining with high levels of accuracy and precision whether a bean is usable or flawed. This process of inspection and classification uses a single algorithm. 

Existing coffee bean data from Central de Cooperativas Agrarias Cafetaleras de los Valles de Sandia, Peru, was used to train the deep learning model topology that makes it easier than ever before to separate the good beans from the bad.  

Lenovo's high-performance computing systems, like ThinkSystem, have helped the complex process speed along.