Abstract:
This study proposes the development of an AI-powered optical mark recognition
(OMR) system that aims to automate and enhance the grading of multiple-choice
question (MCQ) answer sheets. Traditional OMR systems rely on shaded bubbles
or pencil marks, which can cause recognition errors, especially when students use
different marking styles, such as crosses instead of shading. The manual marking
process can take a considerable amount of time for large numbers of students.
To overcome these limitations, the proposed system combines advanced image
processing techniques with a Convolutional Neural Network (CNN) to accurately
classify bubbles as crossed, shaded, or empty. A mathematically calculated and
defined A5 bubble-sheet with a 50-question layout was designed to support two
paper structures, enabling shuffled question papers for enhanced exam security.
These structures make the system more practical, user-friendly, and suitable for
large-scale academic and institutional assessments.
The system was implemented using Python with PySide6 for the graphical
user interface. In addition, OpenCV, TensorFlow, NumPy, and Pandas were
used for processing and model training. A custom dataset was developed using
real student-marked papers representing different marking behaviors and image
conditions. The final CNN model achieved a validation accuracy of 98.87%.
The model demonstrated strong generalization and consistent performance across
different lighting and image quality conditions. Furthermore, the Gemini Pro
API was incorporated to automatically extract student registration numbers from
scanned answer sheets, ensuring accurate mapping between responses and student
identities.
The desktop application provides features for automated scoring, manual verification
by examiners, and exporting results in Excel format. It also supports
evaluating shuffled MCQ papers, enabling examiners to mark answer sheets that
belong to different paper versions with different question orders. Additionally, the
system includes registration number detection. The system substantially reduces
grading time while maintaining human-verified reliability through dual examiner
review features. The solution provides a practical, scalable, and cost-efficient alternative
to commercial OMR systems for large-scale assessments in universities.
