Testing methodology

How We Test Formula OCR

We test formula recognition using a fixed set of printed, handwritten, photographed and scanned documents. Results are evaluated for symbol accuracy, mathematical structure, editability and correction time.

Input coverage

The test set should cover common and difficult documents

Clean printed formulas
Fractions, roots, superscripts and subscripts
Matrices and aligned equations
Mixed text-and-formula documents
Skewed phone photos
Low-resolution images
Handwritten formulas
Chinese or English text mixed with formulas
Scanned PDFs
Failure cases or examples that need visible correction

Evaluation criteria

Results are useful only when limitations are visible

symbol accuracy
mathematical structure
Word equation editability
LaTeX correctness
layout preservation
manual correction time
whether local OCR or optional Cloud AI was used

Required record format

Every public example should show raw output before correction

Source type: Example: phone photo of a printed worksheet
App version: The exact Offline OCR version used
Device: The device or platform used for processing
Processing mode: Local OCR or optional Cloud AI
Expected equation: The target equation or expected document output
Raw OCR output: The unedited recognition result
Corrected output: The result after manual correction
Recognition issues: Known errors or no visible error in this sample
Available exports: DOCX, LaTeX, PDF, text or project file
Processing date: The date the example was processed

We publish raw outputs before manual correction and identify whether each example used Local OCR or optional Cloud AI. Failure cases should remain visible because they help users decide whether a tool fits their documents.