“You need a computer vision system that understands this is a carton even though it might be covered in dirt, even though it might be torn, or even though it might be half-stuck under some other piece of material,” Horowitz said.
His startup now supplies the Alpine Recycling facility in Denver, Colorado with a robotic sorter.
These advanced recycling sorters use the same computer vision honed by companies to power highly automated manufacturing processes, such as building computer chips. However, sorting paper from plastic can be more complicated.
“A human would say ‘these are the types of features I’m looking for’ and then program those into the algorithms to try and find those features,” said Jeff McVeigh, vice president and general manager of visual computing products at Intel. “The computer vision would then make assumptions that this was the type of object that I cared about.”
This identification process works well when materials are extremely orderly and predictable, but the waste in a recycling center is a jumbled, unpredictable mess — even after it has been sorted mechanically.
Identifying a greasy pizza box from a mangled can, neither of which assumes the same orientation or placement on the conveyor belt more than once, requires building out a large representative dataset. Using thousands of images of trash in different positions can help train the neural network, ultimately allowing it to learn on its own, said McVeigh.