Thèse Recherche Rapide de Propriétés d'Images avec Application à l'Artifactométrie H/F

Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Mathématiques
École doctorale : Mathématiques Hadamard
Laboratoire de recherche : Centre Borelli
Direction de la thèse : Miguel COLOM ORCID 0000000326360656
Début de la thèse : 2026-10-01
Date limite de candidature : 2026-09-01T23:59:59

La contrefaçon de produits a des conséquences négatives importantes. En effet, dans l'économie mondiale, des centaines de milliards d'euros sont perdus chaque année à cause de la fraude, ce qui touche une grande variété de secteurs comme l'industrie pharmaceutique, l'électronique, les produits de luxe ou les pièces automobiles, pour n'en citer que quelques-uns. C'est pourquoi il faut des méthodes pour vérifier de manière fiable l'identité des produits et prévenir la fraude tout au long de la chaîne d'approvisionnement. En général, on colle une étiquette avec une sorte d'identifiant visuel (un code de type QR, par exemple) sur le produit. C'est un vrai défi, car les contrefaçons sont de plus en plus sophistiquées chaque jour grâce aux méthodes basées sur l'IA.
Cependant, il est toujours possible de détecter les contrefaçons, car on sait que les systèmes d'impression laissent des traces détectables, différentes de celles du produit d'origine. Ce projet de doctorat explore des stratégies de recherche et d'échantillonnage dans des espaces latents qui ne sont pas explicitement démêlés. Dans de telles représentations, les attributs d'intérêt - comme l'identité du produit, le lot de fabrication, les conditions d'acquisition ou le type de dommage - peuvent être corrélés et difficiles à isoler.
De plus, de nouveaux attributs peuvent ne devenir pertinents qu'une fois la représentation latente établie. Dans ce projet, on va concevoir de nouvelles méthodes permettant de comparer rapidement les traces trouvées dans l'espace latent au sein d'une très grande base de données d'images, ainsi que développer de nouveaux algorithmes et structures de données permettant de récupérer efficacement des exemples similaires possédant un attribut donné, avec pour objectif final de détecter les produits contrefaits à partir de leurs identifiants visuels.

Although copy-detection patterns have been extensively studied for authentication, most existing approaches focus on binary classification (authentic vs. counterfeit). These methods do not address the broader problem of exploring large databases of fine-detail images in order to retrieve examples sharing specific attributes such as acquisition conditions, manufacturing batch, or damage type, with the associated problems such as the need for fast retrieval in a very large database.
At the same time, modern image retrieval systems rely on compact neural embeddings that allow efficient similarity search. However, these embeddings typically encode multiple semantic factors in an entangled latent space, where attributes are not explicitly separated. As a consequence:

- It is difficult to retrieve examples sharing a specific attribute while ignoring other correlated factors.
- Newly relevant attributes may emerge after the latent representation has been learned, making retraining impractical.
- Existing approximate nearest-neighbor methods assume explicit metadata filters, which are often unavailable when attributes are only implicitly encoded in the representation.

Therefore, there is large room for research on how to efficiently search or sample from a fixed, entangled latent representation in order to retrieve examples sharing a newly defined attribute, using only a small set of examples that may be biased.

The objective of this research is to develop methods for efficient exploration and attribute-based retrieval in entangled latent representations of fine-detail images, with applications to copy-detection patterns and scientific image archives. The research will investigate how to retrieve examples sharing a given attribute when this attribute is not explicitly disentangled in the learned representation and may only be defined after the latent space has been constructed.

The objective of this research is to develop methods for efficient exploration and attribute-based retrieval in entangled latent representations of fine-detail images, with applications to copy-detection patterns and scientific image archives. The research will investigate how to retrieve examples sharing a given attribute when this attribute is not explicitly disentangled in the learned representation and may only be defined after the latent space has been constructed.
The proposed methodology combines representation analysis, search structures, and attribute-learning strategies.
First, a baseline latent representation will be obtained using neural encoders trained on image datasets of textures and fine-scale patterns. Existing representation-learning techniques such as self-supervised learning or metric learning will be evaluated to obtain compact embeddings suitable for large-scale indexing.
Second, the structure of the latent space will be analyzed to understand how attributes such as acquisition modality, texture characteristics, or damage patterns are encoded. This analysis will guide the design of search strategies capable of identifying directions or regions of interest within the latent space.
Third, the project will develop efficient search structures enabling attribute-conditioned retrieval. Possible approaches include graph-based indices, tree-based partitions, and multi-index search schemes adapted to latent representations. These structures will allow rapid retrieval of candidate examples in large databases of latent vectors.
Fourth, the project will investigate methods to learn new attributes from a small set of examples. Given a set of query examples representing a newly defined attribute, algorithms will infer a search strategy capable of retrieving additional examples sharing that attribute. Particular attention will be given to the robustness of these methods under dataset bias and correlated latent factors.

The PhD program is organized into four main stages:
1. Representation Learning and Latent Space Analysis
2. Latent Space Exploration and Sampling
3. Efficient Attribute-Aware Retrieval
4. Validation and Application

The PhD candidate will start by reviewing the state of the art and to implement the most relevant methods, to get familiar with both the methods and the data. Then, the results obtained and their research interest will drive on which of the next steps to focus.