Since some dependencies of the service are platform dependent (ARM64 vs X86_64), a small wrapper around the
docker compose
command was created to run the service.
You can use it just like docker compose
:
./passport-service up -d
or
./passport-service up
Once the service is up the API documentation is available at http://localhost:8080/docs.
Place your YOLO model exported as a .onnx
file in the data/models directory.
Make sure it's correctly exported for batch processing (see how to do it here).
To detect passport you first need to place your documents inside the data/passports directory.
To process passport run:
curl -X 'POST' \
'http://localhost:8080/passports/preprocessing-tasks' \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{
"docs": "my_doc_dir",
"detection_args": {"model_path": "model_v0.onnx"},
"batch_size": 64
}'
This will return you something like:
create-preprocessing-tasks-fd05a3b4d2774b269001cccce2fcf073
You can follow the task progress running:
curl -X 'GET' 'http://localhost:8080/tasks/create-preprocessing-tasks-fd05a3b4d2774b269001cccce2fcf073' \
-H 'accept: application/json'
and read the progress from the response.
Then when it's done, you can get the list of preprocessing tasks you can call:
curl -X 'GET' 'http://localhost:8080/tasks/create-preprocessing-tasks-fd05a3b4d2774b269001cccce2fcf073/result' \
-H 'accept: application/json'
Similarly, you can follow each preprocessing task progress.
Preprocessing task will then trigger passport detection on the output of the preprocessing.
You can get the ID of the created detection task from the preprocessing task result:
curl -X 'GET' 'http://localhost:8080/tasks/preprocess-docss-090943423909090000/result' \
-H 'accept: application/json'
When the deteection task is done, you can also call the above API with it ID to get the final output which should look like this:
[
{
"doc_path": "passports/passport.odt",
"doc_pages": [
{
"page": 0,
"passports": [
{
"class_id": "passport",
"confidence": 0.9391622543334961,
"box": [
57.185882568359375,
216.44244384765625,
375.8140869140625,
283.35162353515625
],
"scale": 1.2375,
"mrz": {
"country": "EOL",
"metadata": {
"mrz_type": "TD3",
"valid_score": 62,
"raw_text": "P<EOLSMITH<<JANE<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n01PP300009EOL8107145F2212315<<<<<<<<<<<<<<02",
"type": "P<",
"country": "EOL",
"number": "01PP30000",
"date_of_birth": "810714",
"expiration_date": "221231",
"nationality": "EOL",
"sex": "F",
"names": "JANE",
"surname": "SMITH",
"personal_number": "<<<<<<<<<<<<<<",
"check_number": "9",
"check_date_of_birth": "5",
"check_expiration_date": "5",
"check_composite": "2",
"check_personal_number": "0",
"valid_number": false,
"valid_date_of_birth": true,
"valid_expiration_date": true,
"valid_composite": false,
"valid_personal_number": true,
"method": "direct"
}
}
},
{
"class_id": "passport",
"confidence": 0.9231931567192078,
"box": [
58.3187255859375,
54.883644104003906,
373.454345703125,
185.15440368652344
],
"scale": 1.2375,
"mrz": null
}
]
}
]
},
{
"doc_path": "passports/not_a_passport.jpg",
"doc_pages": []
}
]
Notice that for each document, only pages with passports inside them are reported. On these page, there can be many passport pages, the algorithm outputs a bounding box and optionally a Machine Readable Zone (MRZ) for each one of these pages.
- pipelined pre-processing and detection
To overcome some OpenCV limitations, export the pytorch checkpoint using:
model.export(format='onnx', imgsz=640, dynamic=True, opset=12, verbose=True, simplify=True)
Then, follow the instructions of this issue to make OpenCV work with dynamic batch size.
- simplify using
onnxsim
- set the
height
andwidth
using:
python -m onnxruntime.tools.make_dynamic_shape_fixed --dim_param <param> --dim_value 640 <input> <output>