Computer scientists from the University of Lincoln’s Robotics Lab have built a trainable potato spotting system using low-cost, off-the-shelf hardware.
The Trainable Anomaly Detection and Diagnosis system is able to detect dodgy potatoes, and uses a standard desktop computer along with a low-cost vision sensor to pick out any unwanted spuds. The system also makes use of a GPU of the sort used to help process images in games.
“What a GPU is doing in a games context is turning information into graphics or images, whereas we’re using it in a reverse way to extract information from images," said Dr Tom Duckett of the University of Lincoln.
The system uses state-of-the-art image processing and machine learning techniques to automatically learn the appearance of different defect types. “Existing computer vision systems have to be programmed and calibrated. Our system is different, it learns from samples provided from a human expert," explained Duckett.