Multi-Layer Perceptron-based Classification of Recyclable Plastics from Waste using Hyperspectral Imaging for Robotic Sorting

Abstract: Rapid usage of plastics is a threat to the environment because of their higher persistence in the environment. Automatic Sorting of various kinds of plastics from a waste stream using robot manipulators is preferable as manual sorting is time-consuming and hazardous to the workers involved. We propose the use of hyperspectral imaging with a wider wavelength range in the NIR-zone in conjunction with advanced computing for detecting the recyclable plastics in the mixed waste stream. This paper presents an algorithmic approach based on entropy and contrast stretching for performing image segmentation to obtain pixel clusters representing distinct objects in a given Hyperspectral Image. Obtained pixel clusters with known class labels are used to train a Multi-Layer Perceptron (MLP) neural network model. The trained model is used for the accurate classification (reported 98.2% accuracy on test data) of pixel clusters from unknown waste stream sample images. In the future, we plan to integrate a robotic pick-and-place system for automated sorting and thereafter scale up the developed system to automate the sorting operation in the plastic recycling centers or material recovery facilities (MRF).