Multimodal Classification of PET Waste: An Approach Based on the Fusion of Computer Vision and Spectral Data

Sanchez R.; Flores D.; Becerra J.; Nieves A.; Carrizales J. · 2026 12th International Conference on Mechatronics and Robotics Engineering, ICMRE 2026 (2026) · DOI: 10.1109/icmre69538.2026.11533892

Resumen

The manual sorting process for PET plastics limits the purity of the recovered material, while automated techniques, such as NIR spectroscopy, face restrictions in identifying all plastic categories. This work proposes the design of a system that seeks to overcome these limitations through the fusion of data from a spectral sensor (AS7265x) and images captured by a camera. The system integrates descriptors obtained from $\mathbf{1 8}$ spectral channels in the 410 to 940 nm range, combined with visual descriptors (shape, color, and texture) extracted through digital image analysis, which feed into a deep neural network for final classification. The main objective is to quantify the improvement in precision of the hybrid system compared to unimodal approaches, whose reported performance for classifying PET has been limited. The model was validated experimentally, achieving an overall accuracy of 98 %. Specifically, the system demonstrated a mAP of 97.6% for the PET class, confirming the high viability of this multisensor solution for recycling environments.

Instituciones: Universidad Peruana de Ciencias Aplicadas