Multispectral data classification with deep CNN for plastic bottle sorting

Abstract

Current global trends and green policies indicate the importance of smart waste sorting. Polymer type identification plays a key role in the circular economy model, where high precision is vital to reduce the impurities of recycled plastic flakes. In this paper, we present a robust, high-accuracy plastic bottle polymer type classification using Convolutional Neural Network (CNN). Near-infrared (NIR) absorbance spectroscopy is used to gather polypropylene (PP), polyethene terephthalate (PET), high-density polyethene (HDPE), and low-density polyethene (LDPE) spectra in a dry and wet state. We propose a data augmentation method that generates additional training examples, and we experimentally determine the impact of the ratio of real and generated samples on the accuracy of the classification. In addition, we compare this classification approach with Support Vector Machine (SVM), Principal Component Analysis (PCA) and t-distributed Stochastic Neighbour Embedding (t-SNE) clas-sification methods and also provide data-preprocessing steps for these methods. Finally, we combine pre-processing, component analysis, and CNN to achieve 98.4% accuracy rate while reducing the sizes of CNN input feature vectors and the CNN model itself.