Research on Waste Plastics Classification Method Based on Multi-Scale Feature Fusion

Area of Science:

• Environmental Science
• Computer Vision
• Materials Science

Background:

• Non-degradable plastic waste, particularly from bottles, poses significant environmental challenges.
• Effective recycling necessitates accurate identification and sorting of different plastic types.
• Current identification methods may lack the precision required for complex waste streams.

Purpose of the Study:

• To develop an advanced method for accurately identifying and classifying plastic bottles using multi-sensor data fusion.
• To improve the accuracy of distinguishing between transparent polyethylene terephthalate (PET), blue PET, and transparent polypropylene (PP) bottles.
• To enhance the efficiency of plastic waste recycling processes.

Main Methods:

• A multi-scale feature fusion method, RHFF-SOLOv1 (Segmenting Objects by Locations), was developed.
• Synchronous RGB and near-infrared (NIR) hyperspectral images were acquired using line-scan and hyperspectral cameras.
• A hyperspectral feature band selection method was employed to reduce dimensionality, focusing on the 1087.6 nm to 1285.1 nm range.

Main Results:

• The RHFF-SOLOv1 method demonstrated improved plastic bottle classification accuracy compared to the standard SOLOv1.
• The overall classification accuracy achieved was 95.55%.
• The method achieved a superior accuracy of 97.5% specifically for blue bottle classification, outperforming most other space-spectral fusion methods.

Conclusions:

• The proposed RHFF-SOLOv1 method effectively integrates RGB and hyperspectral data for enhanced plastic bottle identification.
• This advanced technique offers a significant improvement in recycling accuracy, contributing to environmental sustainability.
• The method shows strong potential for practical application in automated waste management systems.