Research on Real-Time Classification Methods for Plastics on Industrial Conveyor Belts Based on Hyperspectral Imaging

To address the real-time requirements in industrial plastic sorting, this study proposes a lightweight hyperspectral classification system scheme for industrial conveyor belt scenarios. Utilizing a high-speed push-broom hyperspectral camera to acquire plastic spectral reflectance curves, an S-G filtering-first derivative-SNV normalization preprocessing pipeline was established. Under the identical preprocessing framework, the classification performance of Partial Least Squares Discriminant Analysis (PLS-DA) and Support Vector Machine (SVM) models was comparatively analyzed on a dataset comprising 24,159 spectral samples from five types of plastics (ABS, PBT, PC, PP, TPU). The results demonstrate that the PLS-DA model achieved a total processing time of 18.18 ms, which is 7.5 times faster than the SVM model's 136.55 ms. The classification accuracy of PLS-DA was 97.33%, exhibiting only a marginal decrease of 0.36% compared to SVM's 97.69%. This validates the application potential of the proposed lightweight scheme for large-scale industrial hyperspectral plastic classification.