PaddleOCR

Engineer Details Six Weeks Spent Training and Testing Signature Redaction Models for Closed-Loop Document Anonymization

A detailed case study describes attempts to automatically redact handwritten signatures from scanned and text-based PDFs containing personal data such as names, phones, addresses and signatures. The work was performed inside an air-gapped environment on a single GPU machine with no internet access. Multiple approaches including color-based ink gates, pre-trained YOLO detectors, custom-trained YOLO11s models, Tesseract OCR heuristics and various vision-language OCR engines were evaluated on 492 real pages plus synthetic augmentations. Key findings include rotation handling bugs, line-assembly failures in Tesseract, the necessity of using apply_redactions instead of draw_rect for true removal, and the limited value of vision models once rule-based pipelines are mature. The final pipeline reduced expensive vision-model calls from 50 pages to 18 pages while achieving zero leaks across thousands of redactions. The author also measured twelve OCR engines and demonstrated that combining PaddleOCR detection with Tesseract recognition yields the best accuracy-to-speed trade-off.