Robust near-infrared-based plastic classification with relative spectral similarity pattern

Area of Science:

• Material science and spectroscopy
• Environmental science and waste management
• Computer vision and machine learning

Background:

• Near-infrared (NIR) hyperspectral imaging offers efficient material recognition.
• High-dimensional spectral data requires effective feature extraction for accurate classification.
• Surface contamination and spectral noise significantly impair material identification accuracy, especially in waste sorting.

Purpose of the Study:

• To develop a robust, real-time feature-extraction method for material classification in noisy environments.
• To improve the performance of hyperspectral imaging in challenging conditions like plastic waste sorting.
• To address the limitations of traditional methods in handling spectral noise and surface contamination.

Main Methods:

• Proposed a novel real-time feature-extraction technique: relative spectral similarity pattern color mapping (RSSPCM).
• RSSPCM compares relative intra- and inter-class spectral similarity patterns, not just individual spectra.
• Utilized intra-class similarity ratios based on similar chemical compositions for feature extraction.

Main Results:

• The RSSPCM method demonstrated robust material classification in noisy environments.
• High accuracy was achieved, with average F1-scores of 0.99 for low-noise and 0.96 for high-noise datasets.
• The method showed minimal F1-score variation across different material classes (standard deviation of 0.026 for high-noise data).

Conclusions:

• RSSPCM effectively overcomes spectral noise and contamination issues in hyperspectral material identification.
• The method provides a reliable solution for real-time material sorting in industrial waste management.
• Relative spectral similarity patterns offer a more robust approach than individual spectral analysis for classification under adverse conditions.