Deep fusion of hyperspectral images to enhance polymers recycling: from a better characterization to cross-modal generative models

Statut

À pourvoir

Disciplines scientifiques

Informatique et Sciences de l'Information

Direction de recherche

Physique et analyse

Site de rattachement

Lyon

The increasing demand for plastics and the need for efficient recycling are critical issues in waste management due to the complexity of separating mixed polymers and impurities. Despite advancements, current recycling rates remain low, largely because of the difficulty in accurately sorting plastic waste. Deep learning and artificial intelligence (AI) are being integrated into sorting technologies to improve the classification and separation of plastics. Among the most used techniques are hyperspectral imaging in the near-infrared range (HSI-NIR) and laser-induced breakdown spectroscopy (LIBS). When combined with machine learning, from chemometrics to deep learning, these methods become powerful tools for sorting plastics. HSI-NIR is effective for identifying the principal molecular components, while LIBS captures elemental signatures. However, each method has limitations, such as challenges with black plastics for HSI-NIR and molecular complexity for LIBS. The fusion of these techniques, enhanced by deep learning, holds the potential to overcome these limitations, improving both the accuracy and speed of sorting processes. The proposed thesis aims to combine HSI-NIR and LIBS imaging technologies, offering better molecular and elemental quantification for plastic sorting, thus improving recycling efficiency. To go further, the student will work on cross-modal generative models to offer an ultimate tool based on spectral analysis and deep learning. Generative deep learning, particularly using autoencoders or diffusion models, is also playing a crucial role by augmenting incomplete datasets. This approach could generate synthetic spectral data, enabling the use of multimodal data for more robust plastic classification.

Keywords: Polymers, sorting, deep learning, generative deep learning, cross-modal modelling, multimodal

  • Academic supervisor    Dr Marion LACOUE-NEGRE, IFPEN, ORCID 0000-0002-1092-2223
  • IFPEN supervisor    Dr Aurélie PIRAYRE, IFPEN, ORCID 0000-0003-0112-3689
  • L@bISEN supervisor    Dr Mohammed El Amine BECHAR, ORCID 0000-0002-9569-357X
  • PhD location    IFPEN, Lyon, France 
  • Duration and start date    3 years, starting in the fourth quarter 2025 (November 12)
  • Employer    IFPEN
  • Academic requirements    University Master degree in mathematics, computational science
  • Language requirements    French, English level B2 (CEFR) (Willingness to learn French)
  • Other requirements    The student must have an open mind for physicochemical analyses and polymer chemistry.

To apply, please send your cover letter and CV to the IFPEN supervisor  indicated here below.

Contact
Encadrant IFPEN :
Dr Marion LACOUE-NEGRE