GUAN Yibo, DU Xinyun, ZHANG Zhixia
The work aims to explore the potential of combining near-infrared (NIR) spectroscopy with a deep learning model for rapid material identification, to address the limitations of traditional detection methods for plastic packaging materials, such as low efficiency, high human dependency, and difficulty in distinguishing visually similar materials with subtle compositional differences. NIR spectral data were collected from representative plastic packaging materials, including PET, PE, PP, PVC, PLA, and PBS. The raw spectra were preprocessed using a multi-step strategy involving Savitzky-Golay smoothing, Standard Normal Variate Transformation (SNV), and Multiplicative Scatter Correction (MSC) to eliminate noise and scattering interference. Based on this, a classification model based on Multi-Scale Mamba (MS-Mamba) was constructed. This model leveraged the linear complexity advantage of the State Space Model (SSM), employing multi-scale convolutional branches and a gated fusion mechanism to simultaneously capture local texture features and global long-range dependencies in spectral sequences. To assess performance, the proposed method was compared with traditional machine learning models (PCA-SVM and PLS-DA) and deep learning models (LSTM and Transformer), with accuracy, precision, recall, and F1-score as evaluation metrics. The experimental results demonstrated that the MS-Mamba model significantly outperformed traditional machine learning (PCA-SVM, PLS-DA) and mainstream deep learning models (ResNet, Transformer) in both classification accuracy and stability. The test set accuracy, precision, recall, and F1 score reached 99.91%, 99.89%, 99.88%, and 99.88%, respectively, while exhibiting exceptional robustness in distinguishing highly similar materials such as PE and PP. The proposed approach enables rapid, non-destructive, and highly accurate identification of plastic packaging materials, providing a feasible and effective solution for automated inspection and sustainable recycling applications.
