Diagram Analysis for Visually Impaired Readers

We present the first complete software-only framework for making UML class diagrams accessible to visually impaired readers. Existing approaches such as tactile graphics, 3D-printed diagrams, or classifier-based models are either costly, domain-limited, or unable to extract full diagram semantics. In contrast, our system combines formal, shallow semantics, YOLOv8-based object detection, dual OCR text extraction, and a structured query interface. The method automatically reconstructs classes, attributes, methods, and relationships, producing XML and accessible text/audio outputs. Evaluation performed on 379 annotated UML diagrams achieved a mean average precision ([email protected]) of 0.910 across 17 element classes, with precision exceeding $95 \%$ for most components. A feasibility study with a blind engineer confirmed the usability of the query-based interface. This work demonstrates that technical diagrams can be fully and reliably transformed into accessible representations without hardware assistance, opening the way for scalable non-visual access to diagrams across software engineering and related fields.