Workflow management tool for making scanned pdf accessible using ocr and ai

Summarise with VIRAT
Authors: Amir Suleimenov
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Scanned PDFs remain a challenging format for digital accessibility. Unlike digitally generated PDFs, scanned documents consist only of images and lack structural tags and searchable text layers, making them unusable for assistive technologies such as screen readers. Existing accessibility tools provide partial solutions but are often fragmented, costly, and dependent on extensive manual intervention. They rarely offer an integrated workflow that combines OCR, structural layout analysis, and automated description generation. This thesis presents the design and implementation of a web-based workflow management system that automates accessibility enhancement for scanned PDFs using vision models and LLMs. The system is implemented with the Django framework and follows a modular architecture that connects AI services with a persistent SQL database. It integrates several key components: DocLayout-YOLO for document layout detection and multimodal language models (LLaVA 1.6 and Gemma 3) for semantic description generation. The platform enables users to upload scanned PDFs, perform layout analysis, extract text through OCR with LLM-based post-processing, and generate page-level or element-level descriptions within a unified interface. Quantitative evaluation shows that the layout detection module achieves an average F1-score of 91.18, demonstrating high precision and recall in identifying document structures. A human evaluation involving five participants, including visually impaired users, indicates that AI-generated descriptions are accurate, fluent, and contextually meaningful. The Gemma 3 model with layout context achieved the highest overall rating (4.82/5) across fluency, correctness, completeness, and specificity. The results demonstrate that integrating OCR, layout recognition, and multimodal AI into a single workflow significantly simplifies the accessibility remediation process.

Fields of Study & Subjects

Digital Accessibility for Disabilities Handwritten Text Recognition Techniques Information Systems and Technology Applications

Publication details

Journal / source
University Library Linz repository (Johannes Kepler Universitat Linz)
Publication date
1 January 2026
Language
EN
Document type
Thesis