École Nationale Supérieure de
Techniques Avancées

Abstract

This internship was carried out in the Multimedia team of the LTCI laboratory at Télécom Paris. Within this team, the PhD student Yasser Benigmim has developed the DATUM method, an unsupervised domain adaptation technique in semantic segmentation using diffusion models. The initial aim of this internship is to adapt this method to image classification. This report therefore contains explanations of the different paths taken to approach this goal, which will not have been achieved as the objective of the internship evolved along the way. The first part of the report deals with the study of the use of customized diffusion models in image classification, while the second focuses on the second objective, which is to improve the Diffusion Classifier method by testing the application of contrastive learning in various forms. Neither objective was reached.