opencr-skill_skill

Overview

This skill extracts text, formulas, tables, and document structure from images, scanned PDFs, and screenshots using OpenOCR. It unifies text detection, recognition, universal VLM recognition, and layout-aware document parsing so you can convert visual content into editable text or structured outputs quickly and reliably.

How this skill works

Initialize the OpenOCR engine with the task you need (det/rec/ocr/unirec/doc) and pass an image, numpy array, or PDF. The engine runs detection and/or recognition models (ONNX or torch), optionally performs layout analysis, and returns bounding boxes, recognized text, confidence scores, or full document blocks. Results can be exported to Markdown, JSON, or visualized for review.

When to use it

• Converting scanned PDFs or images into searchable text
• Extracting mathematical formulas or LaTeX from screenshots
• Detecting and cropping text regions for downstream processing
• Parsing multi-layout documents with mixed text, tables, and images
• Batch processing directories of documents for archiving or indexing

Best practices

• Choose the right task: use 'ocr' for general text, 'unirec' for formulas/tables, 'doc' for full document parsing
• Use mode='mobile' for fast CPU processing and mode='server' when higher accuracy is required
• Prefer the ONNX backend for lightweight CPU deployment; enable GPU via onnxruntime-gpu or torch+CUDA for speed
• Tune thresholds like drop_score (OCR) and layout_threshold (Doc) to balance precision and recall
• Enable layout detection for documents with mixed content and use batch settings (rec_batch_num) to optimize throughput

Example use cases

• Convert a folder of scanned invoices to Markdown and JSON for bookkeeping
• Extract LaTeX from photographed handwritten formulas for research notes
• Detect and crop text regions from photos for an annotation pipeline
• Parse multi-page PDFs with layout analysis and save per-page structured outputs
• Run fast OCR on mobile-collected images for field data capture