PK03 - Automated Grading of Paper Exams using Computer Vision (OMR)

Area: EdTech | Supervisor: Kalchgruber

Problem Statement: Paper-based exams are robust against cheating but impose a massive grading burden on educators. Many solutions (like Moodle) require separate Answer Sheets, which are confusing for students and disconnect answers from the question's context. There is a lack of open-source tools that allow flexible, "single-sheet" exam layouts (where answers are ticked directly next to the question) while still supporting fully automated grading via scanning.

Project Goal: Develop a "Scan-to-Grade" pipeline. The student will build a web application that allows lecturers to (1) Design an exam with auto-generated alignment markers, (2) Bulk-upload scanned PDF submissions, and (3) Automatically grade them using Computer Vision algorithms to detect ticked checkboxes.

Core Tasks:

  1. Exam Generation Engine: Create a module that generates PDF exams with Fiducial Markers (e.g., ArUco markers or QR codes) at the corners. These markers are essential for the software to "understand" the page geometry later.
  2. Computer Vision Pipeline (The Core): Implement a Python backend (OpenCV) that:
    • Detects the corner markers to De-skew and align the scanned image.
    • Applies perspective transformation and high-resolution scaling to flatten the page without pixelation.
    • Isolates and extracts normalized "checkbox regions" based on digital template coordinates.
    • Evaluates checkbox states using both traditional heuristic methods (pixel density, edge detection) and a trained Machine Learning classifier (e.g., SVM), determining whether a box is "Ticked", "Unticked", or "Unsure".
  3. Review Interface: Build a frontend (e.g., React) that shows the scanned exam with the detected answers overlaid, allowing the lecturer to toggle between detection models and manually correct any detection errors.

Research Focus: 
The research focus lies on accuracy evaluation. The student must conduct a stress test using a dataset of ~30 scans with various "messy" marking styles (e.g., faint ticks, crossed-out corrections) to calculate the system's Accuracy, False Positive, and False Negative rated compared to manual grading.

Technologies: React / TypeScript; Python / FastAPI; OpenCV; PDF Generation

Tags: Visual Content Analysis; Multimedia Content Management; Optical Mark Recognition (OMR); EdTech; Automated Assessment; Image Processing; Kalchgruber

Back