Projects

Visual Content Analysis for Reliable Document Access

KATE Graduate School

Funding
Ministry of Science, Research and the Arts Baden-Württemberg.
Role
Doctoral Researcher, KIT CV:HCI.
Host Lab
Computer Vision for Human-Computer Interaction Lab (CV:HCI), Karlsruhe Institute of Technology.
Supervision
Prof. Rainer Stiefelhagen; Prof. Matthias Wölfel.
Current Scope
Document AI, robust visual analysis, accessibility, and reliable multimodal document access.

This doctoral project develops methods that make visually complex documents more reliably accessible. It moves from flexible layout and graph-structured analysis, through robustness under real-world perturbations and domain shifts, toward privacy-aware and semantically grounded document access.

  1. Visual content analysis

    Flexible document layout analysis and graph-based representations of spatial and logical structure.

  2. Robust document analysis

    Evaluation and adaptation under visual perturbations, heterogeneous layouts, and domain shifts.

  3. Reliable document access

    Privacy-aware retrieval and unified parsing of bounding boxes, content, reading order, caption links, and alternative text.

RoDLA robustness benchmark categories for document perturbations
Robust document analysis under layout, content, noise, blur, and consistency perturbations.
GraphDoc relation graph for document structure analysis
Document structure represented through spatial, sequence, parent, and reference relations.

Historical Document Image Analysis

Funding
DAAD IFI: International Research Stays for Computer Scientists.
Role
Visiting Researcher.
Host Lab
Visual Geometry Group (VGG), University of Oxford.
Supervision
Prof. Andrew Zisserman.
Duration
Six-month research visit in 2026.

Conducted within Oxford's Visual Geometry Group under the supervision of Prof. Andrew Zisserman, this visiting project investigates image-based document analysis for historical collections, with an emphasis on visual layout, document structure, and reliable information extraction from heterogeneous archival material.