Federated OMR for Automated Examination Paper Digitization and Assessment
TL;DR
The findings suggest the feasibility of Federated Learning for developing privacy-preserving and highly generalizable OMR systems for large-scale educational assessment.
Abstract
In the context of educational digital transformation, automated grading of paper-based examinations faces two challenges: data privacy concerns and data heterogeneity across educational institutions. Existing Optical Mark Recognition systems are predominantly designed under centralized learning paradigms, requiring examination data to be collected and processed at a central server, which may expose sensitive student information. To address this limitation, this paper proposes a federated OMR framework for automated examination paper digitization and assessment in distributed environments. The proposed system consists of five stages, employing YOLO26 for answer-region localization, ORB-based image registration for geometric alignment, and EfficientNet-B0 for answer-state classification, automated grading, and result aggregation. Federated Averaging is utilized to train a global model across five clients representing heterogeneous data domains with different answer-sheet layouts, ink characteristics, and marking styles, without sharing raw data. Experimental results demonstrate that the federated model achieves an ROI-level accuracy of 99.42% and a Macro F1-score of 95.82%. Furthermore, the proposed system attains a Sheet-level accuracy of 84.62%, outperforming local training approaches while achieving performance comparable to centralized learning. These findings suggest the feasibility of Federated Learning for developing privacy-preserving and highly generalizable OMR systems for large-scale educational assessment.
