The rapid advancement of deep learning-based models has resulted in remarkable successes in supervised learning across various domains, including medical diagnoses, earth observation, and robotics. Machine learning and deep learning techniques are widely applied, even in art conservation, where methods like X-ray fluorescence scanning and hyperspectral imaging are employed to analyze arts-related datasets. It is essential to assess the reliability and effectiveness of these systems before practical implementation, especially considering the uncertainties encountered in real-world scenarios. Neural networks are trained under the assumption that the test data distribution will resemble the training data distribution. However, real-life situations often deviate from this assumption, leading to various errors. Out-of-distribution (OoD) detection aims to prevent these errors by identifying shifts in the test distribution. OoD detection has been well explored in the litterature for multi-class classification but is under-explored for multi-label classification. This thesis aims to fill this gap by providing efficient OoD detection methods and reliability assurances in the context of multi-label classification. This is more challenging than the multi-class context due to complex class boundaries and the need to consider information across different labels jointly. The application of these methods is particularly relevant in non-destructive art investigation and conservation. Imaging modalities like X-ray fluorescence spectroscopy and reflectance hyperspectral imaging provide valuable data for detecting hidden features in artworks, but the analysis of this data requires sophisticated signal processing, image processing, and artificial intelligence techniques. This thesis aims to develop effective solutions for multi-label pigment classification based on X-ray fluorescence or hyperspectral imaging data, providing scientific validity and reliability to the heritage community.
Plus d’informations :
[Site web ADUM]