Deep learning-based plastic classification using spectroscopic data

Plastic recycling represents a significant global challenge. Effective sorting is critical to achieving high-quality recyclate. While infrared (IR) spectroscopy-based classification has become a standard approach, current methods rely on traditional chemometric models and shallow deep learning (DL) architectures, which often struggle with noisy spectral data and limited feature extraction capabilities. To address these challenges, this study proposes a trainable preprocessing module incorporating average pooling layers, layer normalisation, or a combination of both to reduce noise and improve model stability. Additionally, we introduce a Transformer-based deep learning model for efficient plastic classification, alongside improved Convolutional Neural Network (CNN) and Artificial Neural Network (ANN) architectures. Our approach was evaluated on three spectroscopic datasets, including a newly introduced Fourier Transform Mid-Infrared (FTIR) dataset and two open-source near-infrared (NIR) datasets with varying spectral resolutions. The results demonstrate that our proposed methods outperform state-of-the-art machine learning and deep learning models, achieving a 5.4% improvement in F1-score on the FTIR dataset, along with 1.8% and 1.3% improvements in F1-score on NIR datasets, reaching near-perfect classification performance. • A Transformer-based model for classification of plastic using spectroscopy data. • Preprocessing modules to reduce noise and stabilise the model performance. • Improved ANN and CNN models. • An extensive comparative study with state-of-the-art models. • A plastic sorting framework with custom loader, DL models, and training pipeline.