Skip to content
Work

[DOC-01]

DocLayout-YOLO

Fine-tuned YOLOv8 that detects a document page's semantic layout and extracts its text.

  • YOLOv8 (fine-tuned)
  • ONNX Runtime
  • PaddleOCR
  • OpenCV
  • Streamlit

Problem

Raw scanned and digital pages are unstructured. Downstream document AI needs regions, titles, tables, figures, text, cleanly separated before anything else can work.

Architecture

A fine-tuned YOLOv8 model exported to ONNX for cross-platform CPU/CUDA inference detects 11 region classes. PaddleOCR then extracts text per region, and the model outputs structured JSON with bounding boxes, confidence, and text: the clean substrate the rest of a document pipeline can build on.

  1. 01Page imagescanned or digital
  2. 02YOLOv8 → ONNX11 region classes
  3. 03PaddleOCRtext per region
  4. 04Structured JSONboxes · conf · text

Evidence

11
Region classes
detection
~80–96%
Per-region confidence
sample outputs
ONNX
Runtime
CPU / CUDA

The 11 classes: Title, Section-header, Text, Table, List-item, Caption, Page-header, Page-footer, Footnote, Picture, Formula.

DocLayout-YOLO structures a document; Lexora reasons over the structured result and cites its sources. Together they're one document-AI story, raw page to grounded answer.

MIT · v1.0.0