PFCCLab/PPOCRLabel

What it solves

PPOCRLabelv3 is a semi-automatic graphic annotation tool designed specifically for Optical Character Recognition (OCR) tasks. It eliminates the tedious process of manual labeling by using built-in AI models to automatically detect text regions and recognize characters, which can then be manually refined by the user to create high-quality training datasets for OCR models.

How it works

Written in Python3 and PyQT5, the tool integrates PP-OCR models to perform automatic detection and recognition. Users open a directory of images, and the software can automatically generate bounding boxes and text labels. The user then reviews these results, manually correcting text or adjusting box positions using a variety of drawing tools and shortcut keys. Once verified, the labels are exported into formats (like Label.txt and rec_gt.txt) that are directly compatible with the PP-OCR training pipeline.

Who it’s for

It is primarily for developers and data scientists working on OCR projects who need to label large volumes of images for training detection and recognition models, including those working with structured data like tables or Key Information Extraction (KIE) tasks.

Highlights

• Multi-mode Annotation: Supports rectangular boxes, irregular multi-point polygons, table structures, and key information extraction (KIE).
• Semi-Automatic Labeling: Built-in PP-OCR models allow for "Auto recognition" and "Re-recognition" to speed up the labeling process.
• Table Extraction: Includes a specialized workflow to extract table structures and convert them into Excel format.
• Custom Model Support: Users can replace the default ultra-lightweight models with their own custom inference models via command-line parameters.
• Precision Controls: Offers granular control over bounding box vertices using specific shortcut keys (z, x, c, v) for precise alignment.