Towards the Optical Character Recognition of DSLs
summary
Paper evaluates OCR for OCL by generating 3250 images from valid expressions and running Tesseract OCR with four strategies: add DSL punctuation (DP), add DSL lexicon (DPLD), add domain names per expression (DPLD per-expression), and a name-based repairing algorithm post-processor. DPLD per-expression requires retraining per expression but fixes most tokens; the repairing algorithm yields the largest increase in loadable expressions while adding runtime overhead. Handwriting-like fonts drop accuracy significantly.
AI-written summary · about
chapters (8)
- 0:00OCR for DSLs intro
- 0:49How OCR works
- 2:06DSL OCR challenges
- 3:12Improving OCR accuracy
- 5:45Study design and setup
- 7:17Results and analysis
- 8:28Future roadmap
- 9:15Conclusions & Q&A
related talks
40mDeep Learning and Optical Character Recognition
28mSupercharge Information Extraction with Vision-Based LLMs
22mDesign DSLs with xDSL
23mAnyDSL: A Partial Evaluation Framework for Programming High-Performance Libraries
24mDesigning a DSL That Perfectly Solves Your Problem
30mIntroduction to Computer Vision with OpenCV
42mUsing Deep Learning to Understand Documents
22mLex Sleuther - A Novel Approach to Script Language Detection