Optimizing Plastic Identification in E-Waste Recycling through Hyperspectral Imaging and Transformer-Based Machine Learning Models

The rapid growth of the electronics market, driven by high demand for new technologies, has shortened the lifespan of electronic products, leading to a surge in electronic waste (E-waste). Comprising 25% plastics, E-waste contains unrecovered critical and toxic materials, necessitating advanced recycling strategies. HeliosLab, an infrastructure combining imaging sensors and robotic chemical analyses, was developed at the Helmholtz Institute Freiberg. HeliosLab integrates spectroscopy-based modalities such as RGB and hyperspectral imaging (HSI) across multiple wavelength ranges that can be used to optimize E-waste sorting. The complexity of the hyperspectral data, compounded by multisensory integration, requires sophisticated automated algorithms to efficiently process large volumes of data and extract critical material features. These advancements ensure scalable, fast, and automated detection solutions for industrial-scale E-waste recycling operations.

We are developing smart and novel processing methodologies utilizing state-of-the-art (SOTA) machine learning hyperspectral imaging (HSI) classification models. In this study, we focus on Transformer-based architectures, known for their self-attention mechanisms that effectively capture contextual relationships between their input tokens, which enables unique spatial-spectral feature detection, relevant to remote sensing and HSI applications. Such an approach significantly advances automated polymer identification. 

To test the model’s performance on unseen data and evaluate the generalization performance of those SOTA models in industrial-like environments, multiscene datasets are required. We acquired a new multiscene HSI polymer dataset in the near visible (NIR) to the short-wave infrared (SWIR) (400-2500 nm) using hyperspectral cameras available at HeliosLab. The initial deployment highlighted the challenges related to both, the data quality and quantity, as well as regarding methodological frameworks. This led us to develop a tailored Transformer-based topology capable of detecting polymer fingerprints using novel refined extractions of the spatial and spectral features. Our research and advancements contribute to the automation and optimization of polymer detection in E-waste recycling, paving the way for improved resource recovery and environmental sustainability.