Towards Low-cost Plastic Recognition using Machine Learning and Multi-spectra Near-infrared Sensor

This work presents a low-cost sensor and machine learning methods approach for plastic recognition in daily used objects. The sensor is a multi-spectral near-infrared sensor capable of measuring 64 wavelength. Data processing and analysis are performed using a set of four machine learning based computational methods (Random Forest, Support Vector Machines, Multi-Layer Perceptron, Convolutional Neural Networks). Validation is performed by collecting data samples from 6 different types of waste plastics found in household recycling and virgin materials. The results show that Convolutional Neural Networks and Support Vector Machines achieve the highest recognition accuracy of 62.08% with waste plastics and 54.72% with virgin plastics, respectively. The results show how this low-cost multi-spectral near-infrared sensor and machine learning can be effective in plastic recognition tasks and potentially enables to create new applications in other fields that require affordable and portable solutions such as in agriculture, e-waste recycling, healthcare and manufacturing.