OCR Test — Handwritten Math Recognition

Testing local OCR models for recognizing handwritten equations from A-level exam past papers.

Goal

Build a model/pipeline that can read PDFs containing both printed exam questions and handwritten student solutions, without relying on paid API calls.

What We've Tested So Far (on Mac M1 8GB)

Tool	Printed Text	Handwritten Math	Speed	Verdict
pymupdf	Perfect (instant)	Can't read	<1s	Use for printed text extraction
EasyOCR	Decent	Garbage	39s	No good for math
Pix2Tex	Not for full pages	Garbage on handwriting	3-18s	No good for handwriting
Marker	Great quality	Untested (too slow on CPU)	2.3hrs/20pg on CPU	Needs GPU
Nougat	Crashed (dependency issues)	Untested	N/A	Broken deps

Next Step: Test on GPU (RTX 4060)

Marker produced excellent results on printed math (LaTeX output was accurate). Need to test it on the handwritten solutions with a proper GPU.

Setup (Windows + RTX 4060)

1. Install Python 3.11+

Download from python.org if not installed.

2. Create environment

mkdir ocr-test
cd ocr-test
python -m venv venv
venv\Scripts\activate

3. Install PyTorch with CUDA

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

4. Install dependencies

pip install marker-pdf pymupdf easyocr pix2tex Pillow

5. Run the Marker test on handwritten solutions

python test_marker_gpu.py

Sample Papers

• sample_papers/2022 June P1 (Solutions).pdf — Pearson Edexcel IGCSE Further Pure Maths with handwritten solutions in blue ink

Key Files

• test_marker_gpu.py — Main test: runs Marker on the solutions PDF (use this on GPU machine)
• test_pymupdf.py — Quick test: extract embedded text directly (no model needed)
• test_handwriting.py — EasyOCR + Surya comparison test
• test_pix2tex.py — Pix2Tex (LaTeX-OCR) test
• test_ocr.py — Original full comparison script (Marker vs Nougat)

Findings So Far

1. pymupdf extracts all printed text instantly — no model needed for exam questions
2. No off-the-shelf local OCR handles handwritten math — EasyOCR and Pix2Tex both failed
3. Marker is the best candidate but needs GPU (8GB+ VRAM) to run at usable speed
4. If Marker also fails on handwriting, next option is fine-tuning TrOCR on labeled data