Improving Plastic Recycling with Computer Vision

Problem Setting

In most modern plastic waste separation plants, trash gets segregated using high-tech sensors with high detection rates of different plastic types. However, the technology fails on black plastic and plastic that is covered with a “stretch foil” because the plastic has the same black color as the conveyor belt on which the plastic runs, or it cannot be detected since the plastic is hidden behind the foil. This may result in a quality loss of the recycled end product. What is an impossible task for the sensors, we aim to solve using Computer Vision.

Our Solution Approach

We used a transfer learning approach to learn the underlying task with less computational effort using data that we created synthetically. Due to the inexistence of a suitable dataset, we created a synthetic dataset consisting of 40,000 images containing randomly arranged plastic waste of many different kinds, including black plastic and plastic covered by stretch foil. For the model, we chose an SSDLite MobilenetV3 pre-trained on the COCO dataset.


Deployed on a Jetson Nano and with good performance on the synthetic test dataset, as well as promising results in a real test setting, we are optimistic that our computer vision model can support modern plastic waste separation plants to ensure a higher quality of the recycled end product.