But the devices can't identify plastic and other foreign objects that may inadvertently find their way into products during processing.
Such was the motivation for researchers at the Georgia Tech Research Institute (gtri.gatech.edu) to build a computer-vision system that identifies plastic and other unwanted contaminants on a poultry conveyor-line mock-up. A typical poultry production line may move as much as 8,000 lb/hr, says GTRI research engineer John Stewart. That would be a lot of chicken to reprocess or write off if an errant latex glove, for example,-accidentally contaminates the line. GTRI's inspection tool combines computer vision with sophisticated color discrimination algorithms. The system can determine a full range of color, but lab tests have focused on finding blue and green objects. Blue has become a standardized color for plastic used in food-processing. "Few foods are blue, so food processors have hoped that line workers would recognize any foreign objects making their way into the product stream," Stewart explains. "But humans don't make the most consistent inspectors."
Although people are easily trained, they are also easily distracted, adds GTRI research engineer Doug Britton. Product streams typically move at rates of about 12 fps, equivalent to 8 miles/hr. "If a person blinks or looks away for even a second, they can miss a problem," says Britton. "In contrast, machine vision is diligent. It doesn't get tired or bored." What's more, line workers see only the top of finished products. GTRI's computer-vision system captures additional views of surface area by taking digital images as products tumble off one conveyor belt and onto another.
Although such imaging doesn't guarantee the system will spot every single incidence, says Stewart, it should stop subsequent products even if it misses a fragment or two before hand. "The key is to pinpoint where contamination happened and how widespread it is," he says.
The computer-vision system sits above the production line adjacent to conventional metal detectors. The team first trains the system to identify the conveyorbelt background and desired qualities of the food product. This gives a comparison for images captured as products move along the conveyor. If the system sees an object it doesn't recognize, it records the digital image, activates an alarm, and kicks the product off the line.
In lab tests, the system has identified foreign objects as small as 1.5 mm with few false alarms and accuracy rates approaching 100%, researchers report. The system is designed to operate on conveyor belts moving at 12 fps. In the lab, top conveyor speeds were 3 fps. But researchers simulated factory conditions using dim lighting and a longer integration time to produce blur.