Docker Image Repo

https://hub.docker.com/r/spikeyz/mindee-doctr-api

docTR REST API

A Docker image that wraps Mindee's docTR library and exposes all its features through an HTTP REST API built with FastAPI.

Features

• Full end-to-end OCR (image & PDF)
• Text detection only
• Text recognition only (from word crops)
• Key Information Extraction (KIE)
• Export as plain text, hOCR (XML), or searchable PDF
• Annotated visualisation output
• Runtime model selection across 9 detection and 8 recognition architectures
• Model weights pre-downloaded at build time for fast cold starts
• CPU and GPU (CUDA) support

Endpoints

Method	Path	Description
GET	/health	Health check
GET	/models	List available model architectures
POST	/ocr	End-to-end OCR — image or PDF
POST	/detect	Text detection only
POST	/recognize	Text recognition only (cropped word image)
POST	/kie	Key Information Extraction
POST	/export/text	OCR → plain text
POST	/export/hocr	OCR → hOCR XML
POST	/export/searchable-pdf	OCR → searchable PDF
POST	/visualize	OCR → annotated PNG

Interactive documentation is available at http://localhost:8000/docs once the container is running.

Quick start

CPU (default)

docker compose build
docker compose up -d

GPU (NVIDIA)

Requires nvidia-container-toolkit.

Uncomment the args block in docker-compose.yml, then:

docker compose build --build-arg PIP_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cu121
docker compose up -d

Usage examples

# Health check
curl http://localhost:8000/health

# List available models
curl http://localhost:8000/models

# End-to-end OCR on an image
curl -X POST http://localhost:8000/ocr \
  -F "file=@document.jpg"

# End-to-end OCR on a PDF with a different model pair
curl -X POST "http://localhost:8000/ocr?det_arch=fast_base&reco_arch=parseq" \
  -F "file=@document.pdf"

# Detection only
curl -X POST http://localhost:8000/detect \
  -F "file=@document.jpg"

# Recognition only (cropped word image)
curl -X POST http://localhost:8000/recognize \
  -F "file=@word_crop.png"

# Key Information Extraction
curl -X POST http://localhost:8000/kie \
  -F "file=@document.jpg"

# Export as plain text
curl -X POST http://localhost:8000/export/text \
  -F "file=@document.jpg"

# Export as hOCR XML
curl -X POST http://localhost:8000/export/hocr \
  -F "file=@document.pdf" -o output.xml

# Export as searchable PDF
curl -X POST http://localhost:8000/export/searchable-pdf \
  -F "file=@document.pdf" -o searchable.pdf

# Visualise with bounding boxes (returns PNG)
curl -X POST http://localhost:8000/visualize \
  -F "file=@document.jpg" -o annotated.png

Available models

Detection

Architecture	Notes
db_resnet34	Lightweight DBNet
db_resnet50	Default – best accuracy/speed trade-off
db_mobilenet_v3_large	Mobile-friendly
linknet_resnet18	Fast
linknet_resnet34
linknet_resnet50
fast_tiny	Fastest
fast_small
fast_base

Recognition

Architecture	Notes
crnn_vgg16_bn	Default
crnn_mobilenet_v3_small	Lightweight
crnn_mobilenet_v3_large
sar_resnet31
master
vitstr_small	Vision Transformer
vitstr_base
parseq	State-of-the-art accuracy

Configuration

Environment variable	Default	Description
WORKERS	1	Number of uvicorn workers
DOCTR_CACHE_DIR	/app/.cache/doctr	Model weight cache directory

Project structure

.
├── app/
│   └── main.py          # FastAPI application
├── Dockerfile
├── docker-compose.yml
├── .dockerignore
├── .gitignore
└── requirements.txt