Computer Vision | Multimodal AI | Visual Heritage
I build computer vision and multimodal AI systems for visual heritage, image time series, and large-scale visual collections.
Currently at CNRS, I work on AI methods for historical photographs and archival metadata within the Humanum PICTORIA ecosystem, connecting research prototypes with usable datasets and tools.
Research directions
Automatic extraction of visual metadata (text, stamps, handwritten notes) from historical photographs, to help archivists and humanities researchers explore large collections. Release of the FORBIN dataset and detection of stamps to study how photographs circulated.
Features for 2D+t image sequences: a temporal stability measure based on run-length encoding, and Deep-STaR, planar representations that let 2D CNNs learn spatio-temporal patterns. Applied to satellite image time series and violence detection in videos.
Automatic summarization of long videos into short formats for social networks: speech transcription and text clustering, domain adaptation from public benchmarks, and image captioning and tagging with CLIP and BLIP.
Scientific output
Research engineering
A self-attention based object detection model for historical photographs, designed to handle complex visual patterns and diverse object scales.
A collection of 62,135 digitized photographs of Victor Forbin, designed for historical image analysis and archival metadata research.
Photographic colorization service using modern machine learning models based on neural networks. The project was developed during the DataScientest MLOps curriculum.
Spatio-temporal hand-crafted feature extraction for satellite image time series. The proposed method measures stability using a compression algorithm named Run Length Encoding.
A professional LaTeX template for writing a structured PhD thesis, with predefined environments for chapters, sections, figures, tables, citations, appendices and front matter.
Academic and applied path
Courses and resources
Course (in French): Introduction to image processing with OpenCV (C++)
Tutorial: Image classification using deep neural networks
Community
Collaborations
Open to research collaborations and applied computer vision projects.