OCR Vision Workbench
An inspectable OCR workbench for defining regions, tuning preprocessing, and reproducing recognition results.
OCR Vision Workbench turns image text recognition into a visual and reproducible workflow, combining editable regions of interest, configurable preprocessing, Tesseract OCR, result inspection, reusable profiles, and structured exports.
01 — What the Project Does
OCR Vision Workbench is a local, image-first application for configuring, running, inspecting, and reproducing optical character recognition workflows. Instead of treating OCR as a black-box operation that turns an image directly into a string, the project exposes the regions, preprocessing configuration, recognition output, confidence values, and geometry behind each analysis.
Users can load PNG, JPEG, or WebP images, define and edit rectangular regions of interest, navigate the image with zoom and pan, and configure an ordered preprocessing pipeline before running Tesseract-based text recognition.
02 — An Inspectable OCR Workflow
The preprocessing pipeline supports grayscale conversion, resizing, thresholding, and inversion. After an analysis runs, the Workbench preserves more than the final text: results can include word-level tokens, engine-relative confidence, hierarchy information, and bounding-box geometry, allowing recognition output to be inspected in the context of the original image.
Regions and OCR configuration can also be stored in versioned profiles and reused with compatible images. This makes the application useful not only for experimenting with OCR settings, but also for turning a working configuration into a reproducible analysis workflow.
03 — Architecture & Reproducibility
The project is implemented as a local modular monolith. A React and TypeScript WebUI communicates with a FastAPI backend whose application services, core contracts, OCR integration, imaging operations, and filesystem repositories are kept behind explicit boundaries. Tesseract is integrated through an OCR adapter rather than being coupled directly to the core domain.
Geometry is represented explicitly across image, region, crop, and processed coordinate spaces, while the same typed preprocessing model is used by the analysis pipeline. Analyses, reusable profiles, provenance, and structured exports are treated as distinct parts of the workflow so that results can be inspected and reproduced rather than existing only as transient OCR output.
04 — Current Development Status
The core OCR workflow is already implemented, including image intake, region editing, preprocessing, recognition, result inspection, reusable profiles, the REST API, filesystem-backed persistence, and the local WebUI. The project is still in development while regression validation, distribution verification, and the final V1 closure work are completed.
The current distribution target is a local Docker deployment on linux/amd64 using Tesseract for English text recognition. The application is designed for a trusted, single-user environment and currently focuses on OCR of still images rather than PDFs, video, batch processing, or general-purpose computer vision.