Portail HAL des publications du LIRMM
Not a member yet
13279 research outputs found
Sort by
PlantAIM: A new baseline model integrating global attention and local features for enhanced plant disease identification
No full textInternational audiencePlant diseases significantly affect the quality and yield of agricultural production. Conventionally, detection has relied on plant pathologists, but recent advances in deep learning, particularly the Vision Transformer (ViT) and Convolutional Neural Network (CNN), have made it feasible for automated plant disease identification. Despite their prominence, there are still significant gaps in our understanding of how these models differ in feature extraction and representation, particularly in complex multi-crop disease identification tasks. This challenge arises from the simultaneous need to learn crop-specific and disease-specific features for accurate identification of crop species and its associated diseases. To address this, we introduce Plant Disease Global-Local Features Fusion Attention Model (PlantAIM), a new hybrid framework that fuses global attention mechanisms of ViT with local feature extraction capabilities of CNN. PlantAIM aims to improve the model's ability to simultaneously learn and focus on crop-specific and disease-specific features. We conduct extensive evaluations to assess the robustness and generalizability of PlantAIM compared to state-of-the-art (SOTA) models, including scenarios with limited training samples and real-world environmental data. Our results show that PlantAIM achieves superior performance. This research not only deepens our understanding of feature learning for ViT and CNN models, but also sets a new benchmark in the dynamic field of plant disease identification. The code is available at github: PlantAI
GLIMPSE-Med: Single-Screen Visualization of Multivariate Time Series for a Single Individual
No full textInternational audienceThe widespread digitization of hospital information systems is paving the way for the integration of interactive visualization methods into decision support systems. This progress enhances the ability to anticipate critical risks in monitored patients and alleviates the workload of healthcare providers. However, Electronic Health Records (EHRs) encompass large, heterogeneous, and temporal records, making it a significant challenge to develop tools that enable effective understanding trajectories embedded in these complex data. We introduce GLIMPSE-Med, an interactive timeline-based visualization interface for temporal and heterogeneous events in the EHR, incorporating a score generated by a predictive model. The evaluation of this interface, conducted with healthcare professionals, confirmed that it meets two essential needs: (1) Assess the quality of data collected in an EHR ; (2) Estimate the patient's condition over time
D4.2 - FAIR semantic artefact lifecycle from engineering, to sharing
No full textSemantic Artefacts (SAs) is a broad term to designate different knowledge organization systems or semantic resources, including ontologies, terminologies, taxonomies, thesauri, vocabularies and, metadata schemas. SAs are essential for standardising data representation and annotation, encapsulating the highest level of meaningful knowledge within interoperability frameworks. These artefacts are fundamental to sustainable, quality-assured data description practices and they are crucial to follow the FAIR Principles, particularly Interoperability (Principle I.2). Experts, including those involved in the EOSC Interoperability Framework (EOSC-IF), agree that this framework should adopt a bottom-up approach, to take in account insights from different real-world domains. This deliverable (D4.2) presents several aspects of SAs lifecycle that are scanned to guide and inspire stakeholders in producing and distributing FAIR semantic artefacts. D4.2 is one of the main deliverables of FAIR-IMPACT’s task T4.2 which aims to establish guidelines and practices related to the FAIR SA’s lifecycle from creation (Subtask 4.2.1 - Semantic artefact engineering, adoption and description) to sharing and reuse via Semantic Artefact Catalogues (SACs) or repositories (Subtask 4.2.2 - Interoperable Semantic Artefact Catalogues). The task also worked to standardise the mechanisms to describe and serve semantic artefacts (Subtask 4.2.3 - Standardised semantic artefact metadata and their catalogue APIs) respectively with MOD (Metadata Ontology Description and Publication) and the MOD-API specification. This document shows the work of T4.2 in embedding the FAIR principles throughout the lifecycle of SAs (Section 2), from their engineering (Section 2.1) to the sharing process (Section 2.2), with a particular focus on SACs. It highlights the interoperability strategies concerning SACs and the actions taken to federate and sustain them (Section 3)
Parameterizing the quantification of CMSO: model checking on minor-closed graph classes
No full textInternational audienceGiven a graph G and a vertex set X, the annotated treewidth tw(G, X) of X in G is the maximum treewidth of an X-rooted minor of G, i.e., a minor H where the model of each vertex of H contains some vertex of X. That way, tw(G, X) can be seen as a measure of the contribution of X to the tree-decomposability of G. We introduce the logic CMSO/tw as the fragment of monadic second-order logic on graphs obtained by restricting set quantification to sets of bounded annotated treewidth. We prove the following Algorithmic Meta-Theorem (AMT): for every non-trivial minor-closed graph class, model checking for CMSO/tw formulas can be done in quadratic time. Our proof works for the more general CMSO/tw+dp logic, that is CMSO/tw enhanced by disjoint-path predicates. Our AMT can be seen as an extension of Courcelle's theorem to minor-closed graph classes where the bounded-treewidth condition in the input graph is replaced by the bounded-treewidth quantification in the formulas. Our results yield, as special cases, all known AMTs whose combinatorial restriction is non-trivial minor-closedness
Mapping potential environmental impacts of alien species in the face of climate change
No full textInternational audienceRisk maps are a useful tool to prioritise sites for management and allocate resources where they are most needed as they can show us where impacts of biological invasions are most likely to happen or expected to be largest. Given the pace of global changes, we need to understand not only the risks under current conditions, but future risks taking these changes into account. In this study, we use Australian acacias alien to South Africa as a case study to model their potential distribution under future climate change to map their potential impacts at the middle and end of the century and the uncertainty related to three socio-economic pathways and five climatic models. The resulting risk maps across South Africa are a pioneering attempt to combine impacts of alien species with potential future distributions. We found that although climatic suitability and therefore the risk is predicted to decrease under climate change in 51,4% of the country’s area, the opposite is predicted for 26% of the area and the highly vulnerable fynbos biome remains an area with high projected impacts. Such risk maps can help us prioritise management actions and aid the development of suitable plans to protect biodiversity under current and future climate conditions. However, they have to be interpreted with caution and we highlight some shortcomings around species distribution models in general, vulnerability of ecosystems to the potential impacts, data gaps on impacts, as well as currently benign or unknown invaders, which are not included in the projections
Generation and editing of Mandrill Faces: application to sex editing and assessment
No full textVoir aussi le blog de CNRS Le Journal (https://lejournal.cnrs.fr/nos-blogs/focus-sciences/lia-pour-mieux-comprendre-les-mandrills)International audienceGenerative AI has seen major developments in recent years, enhancing the realism of synthetic images, also known as computer-generated images. In addition, generative AI has also made it possible to modify specific image characteristics through image editing. Previous work has developed methods based on generative adversarial networks (GAN) for generating realistic images, in particular faces, but also to modify specific features. However, this work has never been applied to specific animal species. Moreover, the assessment of the results has been generally done subjectively, rather than quantitatively. In this paper, we propose an approach based on methods for generating images of faces of male or female mandrills, a non-human primate. The main novelty of proposed method is the ability to edit their sex by identifying a sex axis in the latent space of a specific GAN. In addition, we have developed an assessment of the sex levels based on statistical features extracted from real image distributions. The experimental results we obtained from a specific database are not only realistic, but also accurate, meeting a need for future work in behavioral experiments with wild mandrills
A new resampling algorithm for particle filters and its application in global localization within symmetric environments
No full textInternational audienceMobile robots are undergoing tremendous development, which makes them employed in many fields. In this area, global localization in symmetric indoor environments is a commonly encountered problem. One of the commonly used algorithms to solve it, is the Adaptive Monte Carlo Localization (AMCL), which is based on the particle filter algorithm. In this paper, we developed a new algorithm for resampling used within the Adaptive Monte Carlo Localization (AMCL) framework that we called Effective Samples Resampling (ESR). The proposed algorithm is based on a deterministic sample selection, and it is thoroughly tested in real time. Using a considerable amount of simulations, the efficacy and robustness of the AMCL using this technique are validated and compared to certain conventional approaches. They are also tested and validated in various real-time operating conditions using the Robot Operating System (ROS). The obtained results are quite satisfying in terms of resampling quality, implementation complexity, and convergence time when compared to random resampling approaches where a sample-based probability density given by high-quality sensors might destabilize localization. The global localization is well handled when the proposed algorithm is involved, compared to standard resampling algorithms that can often be overconfident and fail in some scenarios when there is a lot of symmetry in the considered map of the environment
Fast Autolearning for Multimodal Walking in Humanoid Robots with Variability of Experience
No full textInternational audienceRecent advancements in reinforcement learning (RL) and humanoid robotics are rapidly addressing the challenge of adapting to complex, dynamic environments in real time. This letter introduces a novel approach that integrates two key concepts: experience variability (a criterion for detecting changes in loco-manipulation) and experience accumulation (an efficient method for storing acquired experiences based on a selection criterion). These elements are incorporated into the development of RL agents and humanoid robots, with an emphasis on stability. This focus enhances adaptability and efficiency in unpredictable environments. Our approach enables more sophisticated modeling of such environments, significantly improving the system's ability to adapt to real-world complexities. By combining this method with advanced RL techniques, such as Proximal Policy Optimization (PPO) and Model-Agnostic Meta-Learning (MAML), and incorporating self-learning driven by stability, we improve the system's generalization capabilities. This facilitates rapid learning from novel and previously unseen scenarios. We validate our algorithm through both simulations and real-world experiments on the HRP-4 humanoid robot, utilizing an intrinsically stable model predictive controller
Cartographie multi-échelle des changements dans les récifs tropicaux
No full textThe capacity to observe biodiversity and the marine physical and biological environment is central to support public policies and ensure the ecological and socio-economic sustainability of associated services. In the case of tropical coral ecosystems, observation faces important scientific and technological challenges due to the diversity of these ecosystems, their vast geographical scale and the difficulty to access potentially remote sites. The objective of this PhD thesis is to quantify changes in these ecosystems (surface, bleaching events, species diversity) using a multi-scale automatic image recognition approach (deep learning). This multiscale approach consists in using data collected at fine scale by autonomous boards and sea users (Seatizen project and Ifremer IOT) to identify habitats and species of interest of some characteristic areas of coral ecosystems. The habitats and species identified at this scale will serve as training sets for algorithms using drone data over larger areas. Again, these drone data will serve as training sets for algorithms using satellite images and thus covering very large spatial areas. This approach will be developed at the scale of the Indian Ocean but will be generic to be able to extend the method to other oceans.La capacité d'observation de la biodiversité et l'environnement physique et biologique marin est essentielle pour soutenir les politiques publiques et assurer la durabilité écologique et socio-économique des services associés. Dans le cas des écosystèmes coralliens tropicaux, l'observation fait face à d'importants défis scientifiques et technologiques en raison de la diversité de ces écosystèmes, de leur vaste étendue géographique et de la difficulté d'accéder à des sites potentiellement éloignés. L'objectif de cette thèse est de quantifier les changements dans ces écosystèmes (surface, événements de blanchiment, diversité des espèces) en utilisant une approche de reconnaissance automatique d'images multi-échelle (deep learning). Cette approche multi-échelle consiste à utiliser les données collectées à fine échelle par des planches autonomes et des usagers de la mer (projet Seatizen et Ifremer IOT) pour identifier les habitats et espèces d'intérêt de certaines zones caractéristiques des écosystèmes coralliens. Les habitats et espèces identifiés à cette échelle serviront de jeux d'entraînement pour des algorithmes utilisant des données de drones sur de plus grandes zones. Là encore, ces données de drones serviront de jeux d'entraînement pour des algorithmes utilisant des images satellites et couvrant ainsi de très grandes zones spatiales. Cette approche sera développée à l'échelle de l'océan Indien mais sera générique pour pouvoir étendre la méthode à d'autres océans