Recycling Made More Effective with AI

Canadian recycling facilities reject more than 25% of collected materials. By developing a computer-vision model to sort various materials, we reduced this number, along with the need for human input, and provided quantifiable statistics on the materials passing through a facility.

Challenge

Sorting mixed plastics is one of the first steps in preparing plastic waste for recycling; however, successful separation and identification by plastic types is challenging. In some Canadian jurisdictions, each percentage increase in contamination can further increase costs anywhere from $600,000 to $1 million a year in processing fees and reduce revenue from the sale of recyclables. Recycling plants use systems of computers and machines to autonomously sort glass, plastic, cardboard, and metal. Human sorters must still be utilised to pick out anything the machines missed or mislabeled by hand. It is tedious, repetitive, and expensive work and missed recyclable materials end up adding to costs. Environment and Climate Change Canada (ECCC) requested a proof-of-concept for a solution to more accurately and efficiently sort recyclable materials. Every bit of improvement means fewer recyclables lost to landfills, and less repetitive work for humans, freeing them up to do more valuable tasks.

Solution

We designed an AI model to sort various types of recyclable materials from non-recyclable ones and to differentiate various types of plastic. Our system works in three ways:

  • Using computer vision with cameras positioned over conveyor belts carrying recyclables to feed visual info to our AI model, which uses that data to sort recyclables and pick out garbage.
  • Using semantic segmentation (the ability for the AI to categorize multiple things within one image) to sort contaminants, such as non-recyclable material or another recyclable material that has been mislabeled, from non-contaminants, in this case the specific recyclable material being targeted.
  • Using near infrared spectroscopy (NIR) to identify plastic types — by shooting waves of infrared light at an object you can identify its molecular makeup. Different plastics have different molecular structures, so our solution uses that information to determine which recycling process that piece of plastic requires.
99.5 Accuracy

Results

Not only does our solution sort more efficiently, with accuracy increasing to 99.5%, which decreases the contamination costs incurred, it also offers secondary benefits. It can actively monitor productivity, making note of how much material is being processed at a given time or on a certain conveyor belt, and how accurately that material is being handled, allowing facilities to better understand how to run most effectively. Furthermore, due to freed up people power in their teams efficiency also increased.

Leave a Reply

Your email address will not be published. Required fields are marked *