Robust plastic waste classification using wavelet transform multi-resolution analysis and convolutional neural networks

Mid-infrared spectroscopy (MIR) using photon up-conversion provides advantages over near-infrared spectroscopy (NIR) for plastic waste recycling, including comparable data collection speed and the ability to detect black plastics. However, high-speed MIR spectra suffer from the presence of significant noise. While convolutional neural networks (CNNs) have been utilized for accurate classification of noisy MIR spectra, the analysis of extracted features by the CNN has received less attention. In this study, we analyzed features extracted by a CNN from high-speed MIR spectra collected at 200 spectra per second. Visualizing salient features through the Grad-CAM method revealed that, although the CNN model achieved 100% accuracy, the predictions were not reliable or robust, as the model is susceptible to noise interference. To address this limitation, we propose a wavelet transform-based multi-resolution analysis (MRA) as a preprocessing method for noisy MIR spectra. We show that MRA reconstruction effectively captures features related to characteristic IR peaks, enabling the CNN model to extract informative features from noisy MIR spectra and significantly improves the prediction fidelity and robustness.