Research

My research sits at the intersection of computer vision, human visual attention, and language models. The through-line is perception: how humans look at and understand images, how machines can model that behavior, and how modern AI systems can be made more reliable, efficient, and useful. My PhD (University of Orléans, PRISME Laboratory) focused on modeling attention for paintings and cultural heritage using deep learning, self-supervised learning, and domain adaptation.

Visual attention & saliency

Predicting where and how people look: both the saliency map (where) and the scanpath (the sequence of fixations). This line of work spans natural images, 360° content, and art.

• SATSal: a multi-level self-attention architecture for saliency prediction (IEEE Access).
• SalyPath & SalyPath360: joint saliency and scanpath prediction for standard and omnidirectional images.
• Self-supervised & domain-adaptive scanpaths for paintings, plus the AVAtt art-attention dataset.
• Disc IoU: a perceptually grounded metric for comparing scanpaths, and VLM-based semantic scanpath similarity.

Medical imaging

Deep learning for diagnosis automation and image quality, developed with clinical collaborators.

• Knee-osteoarthritis severity grading with Swin Transformers and graph-prior–aware vision models.
• Deep, domain-adaptive quality assessment of medical images.
• Studying how clinicians look at real vs. AI-generated medical images via eye tracking.

Language models & multimodality

Making language and vision-language models more trustworthy and efficient.

• Classifying and mitigating LLM hallucinations.
• Quality assessment beyond MOS, integrating context, reasoning, and multimodality.
• Efficient small language models and quantization effects on visual perception.

AI agents & AI for science

Retrieval-augmented, tool-using agents and agentic workflows applied to real problems, from a RAG-based coaching assistant to pipelines that accelerate research and evaluation.