HAL - La Rochelle Université, Archives Ouvertes
Not a member yet
19516 research outputs found
Sort by
Extraction d’éléments textuels complexes : Application à la détection et à la reconnaissance des onomatopées dans les bandes dessinées
No full textIn the ever-evolving field of computer vision, scene text detection has emerged as a crucial area of research due to its extensive applications across a wide range of domains. From instant translation to image retrieval, scene analysis, and comprehensive analysis of various documents such as newspapers, historical texts, and notably comic books, the scope of text detection and recognition has significantly expanded. Comic books, a unique blend of art and literature, present a rich tapestry of text styles interspersed with vibrant images and dynamic speech bubbles. While the dialogues enclosed in these bubbles are relatively straightforward to detect and extract due to their uniformity, onomatopoeias present a stark contrast. These artistic interpretations of sound, with their varied complex shapes, orientations, and designs, blend into the graphic elements of the panels, challenging traditional detection methods. Their varied representations further underscore the complexity of designing effective detection techniques. With a growing academic inclination towards the analysis of comic books, especially genres like Franco-Belgian comics, manga, or American comic strips, there is an increased need for robust algorithms. These algorithms must effectively address the natural complexities of these documents. This thesis delves into the world of text detection, leveraging the power of artificial intelligence, with a particular focus on detecting and recognizing complex texts, such as onomatopoeias, in comic books. In this study, an exhaustive literature analysis was conducted on current text detection and recognition techniques. This review aimed to identify the major challenges faced by researchers. Among these challenges are textual variations due to differences in alignment, style, size, and orientation. These variations, coupled with low image contrast and a convoluted background, make automatic text extraction particularly challenging. The current methods, as a result, exhibit relatively low detection and recognition rates, often below 80% and 60% respectively. A new dataset was proposed to address the gaps in character annotation and the presence of complex texts with varied proportions of distortions, alignments, or orientations, including curved text. This dataset's design specifically aims to serve deep learning methodologies requiring a wide range of diversified training data. It encompasses 2,030 images, carefully annotated in characters and words, along with the word class belonging of onomatopoeia based on the source of the emitted sound. It is noteworthy that these images are sourced from the Comicsbook FX website, a platform dedicated exclusively to comic book images segmented into panels. These images were extracted from various comic albums published between 1936 and 1963 by renowned publishing houses Marvel and DC Comics. A significant portion of the Manga109 image base was also annotated for the detection of manga comics. The last part of the thesis presents an analysis of the best approaches to adopt and the proposal of a new approach for the detection and recognition of onomatopoeias, including case examples from Franco-Belgian comics and manga.Dans le domaine en constante évolution de la vision par ordinateur, la détection de texte de scène est devenue un domaine de recherche essentiel, en raison de ses applications étendues dans un large éventail de domaines. De la traduction instantanée à la récupération d'images, en passant par l'analyse de scènes et l'analyse complète de divers documents tels que des journaux, des textes historiques et notamment des bandes dessinées, la portée de la détection et de la reconnaissance de texte s'est considérablement élargie. Les bandes dessinées, un mélange unique d'art et de littérature, présentent une riche tapisserie de styles de texte entrecoupés d'images vibrantes et de bulles dynamiques. Si les dialogues enfermés dans ces ballons sont relativement simples à détecter et à extraire en raison de leur uniformité, les onomatopées contrastent fortement. Ces interprétations artistiques du son, avec leurs formes, orientations et conceptions complexes variées, se fondent dans les éléments graphiques des panneaux, remettant en question les méthodes de détection traditionnelles. Leurs représentations variées soulignent encore la complexité de la conception de techniques de détection efficaces. Avec une inclination académique croissante pour l'analyse des bandes dessinées, en particulier des genres comme la bande dessinée franco-belge, les mangas ou les bandes dessinées américaines, il existe un besoin accru d'algorithmes robustes. Ces algorithmes doivent répondre efficacement aux complexités naturelles de ces documents. Cette thèse approfondit le monde de la détection de texte, en tirant parti de la puissance de l'intelligence artificielle, avec un accent particulier sur la détection et la reconnaissance de textes complexes, tels que les onomatopées, dans les bandes dessinées. Dans le cadre de cette étude, une analyse exhaustive de la littérature a été effectuée concernant les techniques actuelles de détection et de reconnaissance du texte. Cette revue avait pour objectif d'identifier les défis majeurs rencontrés par les chercheurs. Parmi ces défis figurent les variations textuelles dues aux différences d'alignement, de style, de taille, et d'orientation. Ces variations, couplées à un faible contraste d'image et un arrière-plan alambiqué, rendent l'extraction automatique de texte particulièrement ardue. Les méthodes actuelles présentent, de ce fait, des taux de détection et de reconnaissance relativement faibles, souvent en dessous de 80% et 60% respectivement. Un nouvel ensemble de données a été proposé afin de pallier les lacunes en matière d'annotation de caractères et la présence de textes complexes avec des proportions variés de distorsions, d'alignement ou dans diverses orientations, y compris le texte courbé. La conception de cet ensemble vise spécifiquement à servir les méthodologies d'apprentissage profond requérant un large éventail de données d'entraînement diversifiées. Il englobe 2 030 images, soigneusement annotées en caractères et en mots, ainsi que la classe d'appartenance des mots d'onomatopées en fonction de la provenance ou la source du son émis. Il est à noter que ces images proviennent du site web ComicsbookFX, une plateforme dédiée exclusivement aux images de bandes dessinées découpées en cases. Ces images ont été extraites de divers albums de comics publiés entre 1936 et 1963 par les maisons d'édition renommées Marvel et DC Comics. Une bonne partie de la base d'image manga 109 a été annotée aussi pour la détection des comics manga. La dernière partie de la thèse présente une analyse des meilleures approches à adopter et la proposition d'une nouvelle approche de détection et de reconnaissance d'onomatopées incluant des cas d'exemple de comics franco-belge et de manga
Three subspecies of Black-tailed Godwit share non-breeding sites in the world's largest river delta
No full textInternational audienceDuring the non-breeding season (September-April), Black-tailed Godwits (Limosa limosa) are commonly seen in coastal and inland wetlands of the Ganges-Brahmaputra-Meghna Delta in Bangladesh. We hypothesize that the Ganges-Brahmaputra-Meghna Delta, at the overlap between the Central Asian and East Asian-Australasian flyways, may host three subspecies that breed in disjunct areas of temperate and northern Asia: L. l. limosa, L. l. melanuroides, and L. l. bohaii. We used mitochondrial DNA (mtDNA) haplotype network and biometric analysis to determine subspecies in captured individuals, and deployed GPS-GSM transmitters to verify breeding areas of individuals with subspecies assignments. To test for differential habitat preferences, we sampled birds at two ecologically distinct habitats known to host the largest concentrations of non-breeding Black-tailed Godwits in Bangladesh: Nijhum Dweep National Park, a tidal coastal habitat with brackish water on the south-central coast, and Tanguar Haor ('backmarsh'), a seasonal freshwater floodplain in the north. During the non-breeding seasons of 2021-2022 and 2022-2023, we sampled and measured 93 Black-tailed Godwits, 54 of which were equipped with GPS-GSM transmitters. Our mtDNA haplotype network analysis confirmed the presence of limosa, melanuroides, and bohaii subspecies at the study sites. Thus, indeed, Black-tailed Godwits subspecies, despite having distinct breeding ranges, exhibit (partially) overlapping non-breeding ranges in Asia. The subspecies composition differed significantly between sites, with limosa and bohaii dominating in Tanguar Haor and melanuroides in Nijhum Dweep. Of the 21 individuals that were tracked to their breeding grounds, 18 migrated to the expected breeding range of their respective subspecies. However, one bird with a limosa haplotype migrated to a known breeding area of bohaii, whereas two birds with melanuroides haplotypes migrated to the supposed breeding range of limosa. Therefore, while ecological factors at both ends of the flyways may shape the morphological and behavioural differences between Black-tailed Godwit subspecies, their delineations and possible gene flow require further studies
Maximum lifespan and brain size in mammals are associated with gene family size expansion related to immune system functions
No full textInternational audienceAbstract Mammals exhibit an unusual variation in their maximum lifespan potential, measured as the longest recorded longevity of any individual in a species. Evidence suggests that lifespan increases follow expansion in brain size relative to body mass. Here, we found significant gene family size expansions associated with maximum lifespan potential and relative brain size but not in gestation time, age of sexual maturity, and body mass in 46 mammalian species. Extended lifespan is associated with expanding gene families enriched in immune system functions. Our results suggest an association between gene duplication in immune-related gene families and the evolution of longer lifespans in mammals. These findings explore the genomic features linked with the evolution of lifespan in mammals and its association with life story and morphological traits
From Non-overlapping to Overlapping Communities
No full textInternational audienceThis paper presents a new algorithmic framework for transforming non-overlapping community detections into overlapping community structures in networks. Our approach begins with the transformation of the original graph G into a graph G', preparing it for the application of any standard non-overlapping community detection algorithm. After applying a rapid detection of non-overlapping communities in graph G', our method deduces overlapping communities in the original graph G by calculating the degrees of membership for nodes. This method not only captures the complex, multifaceted interactions within networks but also addresses the common question of how to use fast classical community detection algorithms, such as the Louvain method, to identify overlapping communities. Our results on both benchmark and real-world datasets demonstrate the efficiency of our approach in uncovering overlapping community structures
De l'approximation des systèmes dissipatifs au volume élémentaire représentatif en espace-temps.
No full textInternational audienceDe l'approximation des systèmes dissipatifs au volume élémentaire représentatif en espace-temps
Formative Operations in the Critical Zone. Exploration of the Potential Spongescapes of the Zone Atelier Plaine et Val de Sèvre (France).
No full textInternational audienceThe Zone Atelier Plaine et Val de Sèvre is a long-term French research infrastructure specializing in the ecology of cultivated environments. It is located in the north of the Aquitaine sedimentary basin, on the headwaters of the largest wetland on the French Atlantic coast, the Marais Poitevin. Water management in this area has experienced several episodes of crisis over the past three decades, due to the effects of climate change, the development of irrigated crops and excessive drilling, and social contestation of the technical solutions envisaged in response to seasonal deficits, in the form of reservoirs (or “megabassines”). The “Spongescapes” research project is based on the idea of using the ZAPVS perimeter (and the larger Marais poitevin basin) as an experimental site to propose, on the basis of research-creation, prospective representations and workshop tools to help reintegrate nature-based solutions into the animation of local water policies. This poster presents a pannel of graphic representations of potential ZAPVS spongescapes, including elements of field sketchbooks, framework and result of a mapping process, application to agroecological projects
Longevity Collapse in Dolphins: A Growing Conservation Concern in the Bay of Biscay
No full textInternational audienceMarine megafauna populations face global decline from human impacts, making early detection of demographic tipping points essential for effective conservation. Traditional viability analyses rely on Capture-Mark-Recapture data, logistically impractical for highly mobile pelagic cetaceans. Stranding data provide an alternative to conventional monitoring. This study presents the first evidence of declining viability in the most abundant cetacean of the Northeast Atlantic Ocean, the common dolphin (Delphinus delphis) in the Bay of Biscay. Using a novel cross-sectional framework with stratified random sampling, we analyzed age-at-death data from 759 specimens collected between 1997 and 2019. Female longevity declined dramatically from 24 to 17 years, corresponding to a 2.4% reduction in population growth rate. This demographic decline highlights the Bay of Biscay as a demographic sink despite stable abundance estimates. Our findings demonstrate that stranding data can provide acute demographic signals for wide-ranging cetacean species, offering critical early warning indicators for proactive conservation management
Moult is associated with higher diversity of food items in the diet of Common Bulbuls (<i>Pycnonotus barbatus</i>) in Cameroon
No full textInternational audienceMoult is an essential part of birds' annual cycle, and requires sufficient intake of energy and nutrients, but we understand little about how such nutritional requirements are met by wild birds. Using faecal metabarcoding, we analysed the diet of moulting and non-moulting Common Bulbuls Pycnonotus barbatus, captured in Cameroon. We found that the diet of moulting birds was more diverse than that of non-moulting birds, with approximately 1.5 times more arthropod and plant taxa, plus evidence of dietary composition differences between groups. Our results provide novel insight of a likely strategy used by wild birds to fuel the essential self-maintenance task of moult
Numerical Analysis of Energy Retrofits Impacts on Indoor Thermal and Air Quality of Single-Family Houses
No full textInternational audienc
VerifBFL: Leveraging zk-SNARKs for a Verifiable Blockchained Federated Learning
No full textInternational audienceBlockchain-based Federated Learning (BFL) is an emerging decentralized machine learning paradigm that enables model training without relying on a central server. Although some BFL frameworks are considered privacy-preserving, they are still vulnerable to various attacks, including inference and model poisoning. Additionally, most of these solutions employ strong trust assumptions among all participating entities or introduce incentive mechanisms to encourage collaboration, making them susceptible to multiple security flaws. This work presents VerifBFL, a trustless, privacy-preserving, and verifiable federated learning framework that integrates blockchain technology and cryptographic protocols. By employing zero-knowledge Succinct Non-Interactive Argument of Knowledge (zk-SNARKs) and in-crementally verifiable computation (IVC), VerifBFL ensures the verifiability of both local training and aggregation processes. The proofs of training accuracy and aggregation are verified on-chain, guaranteeing the integrity and auditability of each participant's contributions. To protect training data from inference attacks, VerifBFL leverages differential privacy. Finally, to demonstrate the efficiency of the proposed protocols, we built a proof of concept using emerging tools. The results show that generating proofs for local training and aggregation in VerifBFL takes less than 81s and 2s, respectively, while verifying them on-chain takes less than 0.6s