CV / PPOCR

PPOCR

Detects text polygons, rectifies each region, recognizes the text, and returns coordinates and confidence.

7 downloads

Precision

RKNN

Choose the device you're using, the set up guide and documentation will update accordingly.

Platform

Inference Engine

Getting Started

Deploy

sudo docker run --rm --privileged --net=host \
  -e PYTHONUNBUFFERED=1 -e RKNN_LOG_LEVEL=0 \
  -v /proc/device-tree/compatible:/proc/device-tree/compatible:ro \
  ghcr.io/seeed-projects/recomputer-rk-cv/rk3576-ppocr:latest \
  python web_service.py --platform rk3576 --model_dir /app/model --host 0.0.0.0 --port 8000

REST API

Use the REST API to run inference. Copy the commands below.

Curl

curl http://localhost:8080/v1/chat/completions -d '{
  "model": "ppocr-rknn",
  "messages": [{"role": "user", "content": "Hello"}]
}'

Python

import requests

resp = requests.post(
    "http://localhost:8080/v1/chat/completions",
    json={"model": "ppocr-rknn", "messages": [{"role": "user", "content": "Hello"}]},
)
print(resp.json())

Model Details

PPOCR on reComputer RK3576 and RK3588

This end-to-end OCR service comes from reComputer-RK-CV. It detects text polygons, sorts them into reading order, perspective-rectifies each crop, and recognizes every text line.

Model information

Component	File
Text detector	`model/ppocr_det.rknn`
Text recognizer	`model/ppocr_rec.rknn`
Character dictionary	`model/ppocr_keys_v1.txt`
Annotation font	`model/simfang.ttf`
Output	Text, confidence, quadrilateral coordinates, crops, and latency

PPOCR supports still images only. The upstream service intentionally does not include camera, local-video, or MP4-analysis controls.

Run the service

RK3576

bash

sudo docker run --rm --privileged --net=host \
  -e PYTHONUNBUFFERED=1 -e RKNN_LOG_LEVEL=0 \
  -v /proc/device-tree/compatible:/proc/device-tree/compatible:ro \
  ghcr.io/seeed-projects/recomputer-rk-cv/rk3576-ppocr:latest \
  python web_service.py --platform rk3576 --model_dir /app/model \
  --host 0.0.0.0 --port 8000

RK3588

bash

sudo docker run --rm --privileged --net=host \
  -e PYTHONUNBUFFERED=1 -e RKNN_LOG_LEVEL=0 \
  -v /proc/device-tree/compatible:/proc/device-tree/compatible:ro \
  ghcr.io/seeed-projects/recomputer-rk-cv/rk3588-ppocr:latest \
  python web_service.py --platform rk3588 --model_dir /app/model \
  --host 0.0.0.0 --port 8000

Open http://<BOARD_IP>:8000 or /docs.

Startup arguments

Argument	Default	Description
`--platform`	Required	`rk3576` or `rk3588`.
`--model_dir`	`model`	Models, dictionary, font, and sample image.
`--image`	None	Still image analyzed before server startup.
`--output`	None	Annotated JPG/PNG path for `--image`.
`--no_server`	Off	Exit after processing `--image`.
`--host` / `--port`	`0.0.0.0` / `8000`	Service address and port.

To analyze one mounted image, save the annotation, print JSON, and exit:

bash

python web_service.py --platform rk3576 --model_dir model \
  --image /data/document.jpg --output /data/document_result.jpg --no_server

REST API

Endpoint: POST /api/models/ppocr/predict

bash

curl -X POST "http://<BOARD_IP>:8000/api/models/ppocr/predict" \
  -F "file=@document.jpg" -F "drop_score=0.5" \
  -F "include_crops=false"

Optional fields are det_threshold, box_threshold, unclip_ratio, drop_score, max_results, and include_crops. Other endpoints provide health state, configuration, the latest structured result, the latest annotated JPEG, and interactive OpenAPI documentation.

Field	Default	Description
`det_threshold`	`0.3`	Pixel-map detection threshold from 0 to 1.
`box_threshold`	`0.6`	Minimum DB text-box score.
`unclip_ratio`	`1.5`	Polygon expansion ratio from 0.1 to 5.
`drop_score`	`0.5`	Minimum recognition confidence included in combined text.
`max_results`	`100`	Maximum regions sent to recognition.
`include_crops`	`false`	Include base64 JPEG crops in result lines.

Important response fields:

json

{
  "success": true,
  "model": "ppocr",
  "result": {
    "text": "Recognized first line\nRecognized second line",
    "lines": [
      {
        "index": 1,
        "text": "Recognized first line",
        "confidence": 0.9621,
        "accepted": true,
        "box": [[42, 31], [286, 29], [287, 72], [43, 74]]
      }
    ],
    "count": 2,
    "timing_ms": {"detection": 54.2, "recognition": 71.4, "total": 129.8}
  }
}

Detection quality directly affects recognition. Very small, blurred, curved, vertical, low-contrast, or heavily rotated text may be missed, and the simple reading-order heuristic may not suit complex multi-column documents. include_crops=true increases the response size and is intended mainly for the Web results panel. The detector and recognizer are pipeline stages and must remain compatible with the character dictionary and perspective-crop logic.

Build locally

bash

docker build -f docker/rk3576/ppocr.dockerfile \
  -t rk3576-ppocr:local src/rk3576_ppocr

docker build -f docker/rk3588/ppocr.dockerfile \
  -t rk3588-ppocr:local src/rk3588_ppocr

Inputs and Outputs

Input: still image. Output: ordered text lines, confidence scores, quadrilaterals, and annotated image.