Home/Work/№ 08

08 · AI · Vision & OCR

Document Preprocessing & OCR Toolkit

The data-quality layer beneath the document platform: a data-centric story where accuracy came from preprocessing rather than bigger prompts.

PeriodNov 2025StatusProductionAreaAI · Vision & OCR

Highlights

  1. doc_sorter (local VLM categoriser) bootstrapped labels → doc_classifier (ResNet-18) deployed in the hot path: teacher/student at micro scale.
  2. Custom shrinking-rectangle whitespace cropper and background/page-detection cropper; Real-ESRGAN handwriting upscaler with tiled inference.
  3. TrOCR auto-tiling harness, deepseek-ocr and Tesseract probes: documented, reproducible OCR comparisons.