Chafic ABOU AKAR : "Efficient image generation models and strategies for smart industries"

Abstract : Smart robots powered by Deep Learning (DL) models, require extensive annotated data to properly assist human workers. While synthetic data tackles the limitations of real data acquisition, a gap persists between both data domains. Additionally, due to the specificity of the industrial field, there is a need to explore new practices of synthesizing and using industrial image data.
In this thesis, we propose a modular synthetic data generation pipeline to build and maintain large-scale complex environments. Therefore, we render realistic and annotated rich images, and subsequently, introduce the SORDI.ai dataset. We bridged the gap by enhancing the visual properties of 3D models, and enriching the simulation scenes with structured Domain Randomization (DR). However, after training object detection models, we notice the significance of using multi-domain data, to enhance DL generalization, especially in low real data regime. Furthermore, similarly to the structured DR components, we endorse the capability of prompt engineering to ensure the structured variability of AI-generated images. Hence, we propose a multi-source strategy to use simulation and AI-generated data to approach the highest precision when training solely with synthetic data. Then in a second part, we discuss common generative AI methods, applications and modes. For instance, preserving fine-grained and small features of complex images is a common challenge. Therefore, we apply a conditional training after transforming a complex crowded dataset into a larger, and a simpler image dataset. Taken together, we draw best practices for industrial synthetic data.

Jury Composition :

Abdallah MAKHOUL, Professeur, Université Bourgogne Franche-Comté, PhD Director

Cédric  WEMMERT, Professeur, Université de Strasbourg, Reporter

Stéphane CHRETIEN, Professeur, Université Lyon 2, Reporter

Andre LUCKOW, Maître de conférences, Université de Munich, Reviewer

Joe TEKLI, Professeur associé, Université Libano-américaine, Reviewer

Elias DOUMITH, Professeur associé, Université Antonine Liban, PhD Co-director

Localization : Université de Franche-Comté UFR STGI 4 Place Lucien Tharradin 25200 MONTBELIARD - salle Amphithéâtre B