Developing ‘machine vision’ approaches could help producers of insects for alternative feeds improve quality control and quickly identify health problems in their stocks.
Researchers led by Mubeen Tayyab at the University of the West of England have been training and testing an algorithm known as You Only Look Once (YOLO) to identify mealworm beetles and detect anomalies among populations.
The scientists found that when shown images of 50-200 beetles in a tray, precision in identifying individuals was 94%. However, this accuracy dropped to 74% with 500-1,000 individuals.
They found that a solution was to train the model by dividing the image of the larger number of insects into smaller patches. Using this approach, accuracy was restored to 93%.
Insight into feed insect health
The data set used in the study was collected at a mealworm production facility using a technology called Insecto, which records images and video as well as collecting data on parameters such as temperature and humidity.
Ultimately, the automated visual analysis approach will be used to count and size insects as they move through life stages, from larvae to adult. It will also help in detecting anomalies in populations, which may provide early insight into the development of disease issues.
“Such automation will not only help in remote monitoring but in reducing manual labour, costs and improving scalability. Thus, ensuring consistency in quality and quantity of insect production,” wrote Tayyab, in information presented to the Royal Entomological Society’s special interest group on Insects as Food and Feed.
Further work as part of the project will test other object detection models, and trial these approaches using insect larvae as well as adults.
