ED Sciences Chimiques
Formulation and aeration of cereal doughs by means of bio sourced components
by Lucie PERIE (Institut de Chimie & de Biologie des Membranes & des Nano-objets)
The defense will take place at 14h00 - Amphithéâtre B18N Allée Geoffroy Saint Hilaire, Bâtiment B18 33600 Pessac
in front of the jury composed of
- Fernando LEAL-CALDERON - Professeur - Bordeaux INP - Directeur de these
- Romain KAPEL - Professeur - Université de Lorraine - Rapporteur
- Dominique CHAMPION - Professeure - Institut Agro Dijon - Université Bourgogne Franche Comté - Rapporteur
- Véronique BOSC - Maîtresse de conférences - Agroparistech - Examinateur
- Marc ANTON - Directeur de recherche - INRAE Nantes - Examinateur
- Claudia NIOI - Maîtresse de conférences - ISVV - Université de Bordeaux - Examinateur
- Sophie LECOMTE - Directrice de recherche - CNRS Bordeaux - Examinateur
The replacement of controversial additives in cereal matrices represents a major challenge to meet consumers' expectations. Leavening agents are functional ingredients that are required to obtain porous biscuit products according to industrial manufacturing methods. Their incorporation into biscuit dough determines the expansion of dough pieces during the baking stage. In this work, we considered two cereal doughs with different hydration levels that determine the gas incorporation pathways, aiming to completely suppress the need for leavening agents. In a low-hydration laminated biscuit dough, the study considered the use of baker's yeast as a substitute for leavening agents. The configuration of the gluten network conditions the dough elasticity and its ability to stretch to allow the biscuits to expand during baking. In a sponge drop (whipped) dough, air incorporation relies on the formation of a stable foam simultaneously with the gas release induced by the leavening agents. The removal of leavening agents from this matrix was enabled by using functionalized plant proteins through various treatments (physical or enzymatic). A design of experiments approach was implemented to optimize functionalities and thus, ensure the obtention of biscuits with a uniform crumb structure. During this process, the interfacial properties of the dough proteins determine their ability to stabilize the air bubbles in the matrix. These were studied using tensiometry and interfacial rheology.
ED Sciences de la Vie et de la Santé
Computer based analysis of Biological Images Neuronal Networks for Image Processing
by Dirk HILLMER (BoRdeaux Institute of onCology)
The defense will take place at 17h00 - Visio conférence
in front of the jury composed of
- Martin HAGEDORN - Professeur - Université de Bordeaux - Directeur de these
- Christophe MULLE - Directeur de recherche - Bordeaux Neurocampus - Examinateur
- Gertraud OREND - Directrice de recherche - Inserm Universite de Strasbourg - Rapporteur
- Jörg WILTING - Professeur - Universität Göttingen - Rapporteur
AI in medicine is a rapidly growing field, and its significance in dermatology is increasingly pronounced. Advancements in neural networks, accelerated by powerful GPUs, have catalyzed the development of AI systems for skin disorder analysis. This study presents a novel approach that harnesses computer graphics techniques to create AI networks tailored to skin disorders. The synergy of these techniques not only generates training data but also optimizes image manipulation for enhanced processing. Vitiligo, a common depigmenting skin disorder, serves as a poignant case study. The evolution of targeted therapies underscores the necessity for precise assessment of the affected surface area. However, traditional evaluation methods are time-intensive and prone to inter- and intra-rater variability. In response, this research endeavors to construct an artificial intelligence (AI) system capable of objectively quantifying facial vitiligo severity.The AI model's training and validation leveraged a dataset of one hundred facial vitiligo images. Subsequently, an independent dataset of sixty-nine facial vitiligo images was used for final evaluation. The scores assigned by three expert physicians were compared with both inter- and intra-rater performances, as well as the AI's assessments. Impressively, the AI model achieved a remarkable accuracy of 93%, demonstrating its efficacy in quantifying facial vitiligo severity. The outcomes highlighted substantial concordance between AI-generated scores and those provided by human raters.Expanding beyond facial vitiligo, this model's utility in analyzing full-body images and images from various angles emerged as a promising avenue for exploration. Integrating these images into a comprehensive representation could offer insights into vitiligo's progression over time, thereby enhancing clinical diagnosis and research outcomes. While the journey has been fruitful, certain aspects of the research encountered roadblocks due to insufficient image and data resources. An exploration into analysis of in vivo mouse models and analysing pigmentation of skin cells in a preclinical embryo models as well as retina image recognition was regrettably halted. Nevertheless, these challenges illuminate the dynamic nature of research and underscore the importance of adaptability in navigating unforeseen obstacles. In conclusion, this study showcases the potential of AI to revolutionize dermatological assessment. By providing an objective evaluation of facial vitiligo severity, the proposed AI model offers a valuable adjunct to human assessment in both clinical practice and research settings. The ongoing pursuit of integrating AI into the analysis of diverse image datasets holds promise for broader applications in dermatology and beyond.
ED Sciences Physiques et de l'Ingénieur
Definition of an object localization algorithm based on video streams from multiple cameras and evaluation on an MPSoC FPGA
by Denis SHEMONAEV (Laboratoire de l'Intégration du Matériau au Système)
The defense will take place at 14h00 - Amphitéatre JP Dom, IMS - Laboratoire de l'Intégration du Matériau au Système, 351 Cours de la Libération, 33405 Talence Cedex
in front of the jury composed of
- Christophe JéGO - Professeur des universités - Bordeaux INP - Directeur de these
- Daniel CHILLET - Professeur des universités - Université de Rennes - Examinateur
- Anthony BESSEAU - Directeur général - EMG2 - Examinateur
- Matthieu ARZEL - Professeur - IMT Atlantique - Rapporteur
- Andrea PINNA - Professeur des universités - Sorbonne Université - Rapporteur
The recent rise of Convolutional Neural Networks (CNNs) has enabled significant advances in the extraction of visual information by computers. In particular, multi-object detection and tracking are highly required tasks in industry, for applications such as defect inspection or inventory management. However, achieving real-time performance requires significant computational power. Cloud computing partially addresses this need by offloading calculations from the sensor to dedicated servers equipped with powerful and energy-hungry GPUs. However, this approach has several drawbacks, especially for time-critical applications or those requiring data confidentiality. Therefore, at the dawn of Industry 4.0, actors are turning to edge computing. This paradigm advocates allocating computational resources as close as possible to the sensor. Hardware platforms such as FPGA MPSoCs stand out for their flexibility and computational efficiency. However, the methodologies proposed by manufacturers (AMD, Intel) for integrating computer vision functions into these platforms are recent and complex to implement. In particular, there is no methodology for developing applications that exploit multiple cameras. The thesis work presented in this manuscript focuses on the development of an object localization algorithm from multiple video streams for integration into a NATvision platform. This is an FPGA MPSoC platform developed by the company NAT. The first contribution of this study is the definition of a modular algorithm for processing and fusing information from multiple video streams. Special attention has been paid to the execution of the algorithm to simplify its use and the integration of new developments. The prototype was evaluated on a laptop with three video streams passing through an object detector trained on a dataset reduced to three classes. The second contribution concerns the selection, simplification, and implementation of a multi-object tracking algorithm on an FPGA MPSoC. Experiments were conducted to accelerate the SORT algorithm on an FPGA MPSoC. A comparative study of optimized software implementations (SIMD) and hardware implementations obtained using an HLS synthesis tool was conducted. In this context, a hardware accelerator for solving linear sum assignment problems was implemented and evaluated on an FPGA MPSoC. This ultimately enabled the implementation of a complete chain of object detection and tracking on a prorotype platform based on an FPGA MPSoC. After evaluation on a MOT15 dataset, composed of different scenes from the MOTChallenge benchmark, this implementation achieves real-time performance (>15 FPS) while maintaining accuracy close to a GPU implementation.