Honey is one of the world’s most adulterated or mislabeled foods. Last year, Capilano, Australia’s largest honey producer was accused of selling honey that was a mix of cheaper ingredients like cane sugar, corn, and rice syrups. Occasionally fake honey is even unsafe to eat. In 2011, an investigation found that American grocery store shelves were selling honey laced with animal antibiotics, lead, and other heavy metals.
Enter machine learning, which excels at classifying things that humans have a hard time telling apart. The current methods of melissopalynology, or the study of pollen, are often expensive and time-consuming, conducted in laboratories by specialists who require special equipment.
A new, computer vision-powered method that would be much more affordable is laid out in “Honey Authentication with Machine Learning Augmented Bright-Field Microscopy.” The paper was accepted at the “AI for Social Good” workshop at last year’s Neural Information Processing Systems (NeurIPS), a prestigious AI conference.
The team—made up of two college students and a high-schooler—developed the honey authentication tool using a $130 microscope that they found was easy enough for an 11-year-old to use.
“We thus reckon that it would, in practice, prove scalable as a decentralized system where producers/consumers/beekeeping associations are able to test honey easily and help weed out fraudsters,” said Peter He, one of the authors of the paper and a student at Imperial College London, in an email.
They collected different kinds of honey (manuka, acacia, “Lithuanian,” “Black Forest,” eucalyptus melliodora, and thyme) and spread them thinly across glass slides and took images. Approximately 2,500 individual pieces of pollen were annotated and labeled into three categories—round, triangular, and spiky. Pollen in different honey is unique enough that it’s easy for a machine to recognize a valuable manuka honey from one coming from Lithuania.
The system works in two parts. First, a pollen identification neural network detects and identifies the botanical origin, density, and distribution of present pollen. This data is then used in the authentication network, which decides whether the honey is what it says it is.
“Honey samples diluted with sugar syrup can be detected from pollen density analysis, and honey samples diluted with cheaper honeys can be detected from pollen distribution comparison,” the researchers wrote in the paper. “Mislabelled honey samples can be identified through the botanical sources of their pollen.”
The paper focuses on a very small group of 10 samples, which were all classified correctly. Their method, however, is unable to spot contamination with heavy metals, pesticides, or antibiotics or to detect ultra-filtered honey samples that have had the pollen removed, although that in itself is often a sign of adulteration.
Still, the researchers say their approach could make honey inspection cheaper and faster. With a more extensive collection of honey and further research, He said they could create honey “fingerprints” for specific producers and that their approach could be useful for forensic palynology, or the study of pollen and spores for legal matters. Now, the world just needs a way to save bees from colony collapse disorder so there is honey to look at.