Public source, setup notes, benchmark definitions, and roadmap.
Open ↗Project / Evidence-aware document processing
OCR97
A local document and image-text processing tool for screenshots, scans, forms, and image-based PDFs with adaptive preprocessing, lane selection, uncertainty visibility, and inspectable output.
01 / Project
OCR97 treats extracted text as evidence with provenance and uncertainty rather than as a guaranteed transcription ready for silent downstream use.
Status
Alpha / experimental · v0.1.0-alpha · 2026
Technical focus
- Native PDF text and local OCR routing
- Adaptive image preprocessing
- Screenshots, scans, forms, and image-based PDFs
- Confidence and uncertainty visibility
- Source preservation and downstream verification
02 / Design review
Capability is presented with its evidence and boundary.
This record distinguishes what the current project can demonstrate from what remains experimental or unproven.
Problem
- Document automation often passes OCR output downstream without preserving the source image, routing decision, or uncertainty that produced it.
- OCR97 is designed to keep those decisions inspectable so a workflow can distinguish usable extraction from partial or failed evidence.
Design approach
- The pipeline selects between native PDF text and local OCR engines, applies preprocessing where appropriate, and retains source evidence with the result.
- Routing and scoring expose which extraction lane was used and where confidence is limited instead of presenting every output as equally reliable.
- CLI and gateway interfaces support repeatable local workflows without requiring document content to be sent to a hosted OCR service.
Boundaries
- OCR97 does not claim universal character or document accuracy.
- Confidence scores and benchmark results are evidence about a specific dataset, configuration, and scoring method—not proof for every document type.
- Low-quality scans and handwriting require review; downstream systems should not treat uncertain text as authoritative.
Evidence / verification
- A corrected public SROIE receipt-field run scored 92/100 across the first 50 evaluated receipts using the project-specific field metric.
- That bounded result trailed the referenced docTR result of 96.5 and exceeded the referenced Tesseract result of 80 under the documented comparison; it is not a universal accuracy percentage.
- A two-line IAM handwriting check scored 28.5/100, a deliberately visible weak result from a sample too small for broad conclusions.
Current limitations
- Handwriting performance is weak in the current evidence and needs broader evaluation.
- The published samples are bounded and do not establish production reliability across arbitrary document classes.
- Local preprocessing and OCR can have high latency, especially on difficult or multi-page inputs.
03 / Selected artifacts
Source, notes, and inspectable records.
Selected references connect the project description to its public source, technical notes, and evidence records.
Dataset scope, scoring method, and bounded comparison results.
Open ↗Extraction results retain routing, evidence, and uncertainty for downstream review.