A neural network trained on a dataset of images can identify recyclable objects with more than 95 per cent accuracy using a standard webcam.
Recyclable objects ending up in landfill is a major issue for the planet. plastic objects don’t end up being recycled, according to one analysis. Part of the issue is human uncertainty about what items can be recycled.
Being able to rapidly identify what objects are made of, and therefore their recyclability, could help us more effectively decide which bin to put an item in. Now a – a type of neural network often used to analyse visual information – trained on a database of images of rubbish could assist this process.
Advertisement
Developed by Ryan Grammenos and Youpeng Yu at University College London, this network is able to make connections between how an object looks and the materials it is made from. “We’re identifying the material,” says Grammenos.
By deploying extra tricks – including not training certain layers of the neural network and instead devoting computer power to trickier tasks, such as how to interpret malformed or misshapen objects – the researchers were able to further improve the software’s ability to classify materials correctly. The neural network gave the right answer in 95.4 per cent of cases.
“Computer vision researchers often talk about ‘in-the-wild’ recognition challenges, and waste classification is an outstanding example of this,” says at the University of Copenhagen in Denmark. “This study shows promising results using some state-of-the-art methods, and I can see this kind of technology gaining traction in a variety of public settings.”
One of Grammenos’s colleagues at his laboratory is already working on a practical use for the model, having developed a similar system that requires less computational power to classify an object’s recyclability. The goal is to send a signal to raise the lid of the correct bin for any object within 50 milliseconds.
Reference:
