Annonce

M2 (+phD) in Deep Learning for Archeology

13 Février 2024

Catégorie : Stagiaire

Supervised classification of wood and charcoal from microscopic images

M1/M2 internship proposal + possible PhD thesis

Supervisors: Marco Corneli (CEPAM, INRIA, UniCA), Diane Lingrand (I3S, UCA), Frederic Precioso (I3S, UniCA),

Isabelle Thery (CEPAM, UniCA),

Laboratories: I3S, INRIA-MAASAI and CEPAM, UniCA.

Scientific project

Within the framework of the recently funded ANR project AI-WOOD, researchers at CEPAM, I3S and INRIA Université Côte d’Azur are collaborating to the development of new machine/deep learning approaches aiming at performing the taxonomical identification (i.e. classification at the species, genus or family level) of wood and charcoal from microscopic 2D images. The project has a main interest from an archaeological point of view, the main idea being to train a classifier on a modern collection (about 6000 images for 120 species) and then use it to identify ancient charcoals. The anthracologists (i.e. the archaeologists specialized in the identification and analysis of ancient charcoal) actually perform this identification relying on comparative anatomy and based on anatomical features settled by the IAWA1 that they build manually through microscopic observation. Apart from being long and tedious, this identification routine is not entirely satisfying, (also) due to the anatomical proximity of some essences.

Hence, the aim of this project is to explore the potential of machine/deep learning to directly identify the taxon of a specimen from the microscopic image and possibly to boost the identification routine. Although some attempts in this direction have been made in the literature (Rosa da Silva et al., 2022; Silva et al., 2022) there is still considerableroom for improvement.

In this internship we plan to explore several avenues:

1. conceive multiview classifiers (Wang et al., 2020) since each observation/specimen has several “faces”, corresponding to different acquisition zooms and three sections (cross, longitudinal and radial);
2. exploit the IAWA features (available for the training dataset) in order to merge learning approaches and symbolic AI techniques (see e.g. Diligenti et al., 2017; Ciravegna et al., 2023). This might be helpful in order to reduce the number of observations needed to train the model;
3. develop ad-hoc solution (i.e. dedicated convolutional layers) accounting for the specific nature of the data

Candidate profile, timeline and stipend