Artefact detection

The doctr.contrib module gathers complementary features that are useful in document pipelines but are not part of the core OCR. Its first member, ArtefactDetector, finds non-textual elements on a page: bar codes, QR codes, logos and photos.

Note

Contrib modules run on ONNX Runtime instead of PyTorch. Install the extra first:

pip install "python-doctr[contrib]"
# or, for GPU inference: pip install onnxruntime-gpu

Detect artefacts

from doctr.contrib import ArtefactDetector
from doctr.io import DocumentFile

pages = DocumentFile.from_images(["path/to/page.jpg"])

detector = ArtefactDetector(batch_size=2, conf_threshold=0.5, iou_threshold=0.5)
results = detector(pages)

for artefact in results[0]:  # one list per page
    print(artefact["label"], artefact["confidence"], artefact["box"])
# e.g. qr_code 0.91 [812, 64, 1003, 255]

detector.show()  # requires the `viz` extra

Each artefact is a dict with a label, a confidence and a box as [xmin, ymin, xmax, ymax] in pixels.

Parameter	Default	Description
batch_size	2	Number of pages processed at once
conf_threshold	0.5	Minimum confidence of a detection
iou_threshold	0.5	Overlap above which duplicate detections are merged (NMS)

Use your own YOLOv8 model

The detector is based on YOLOv8. You can run any YOLOv8 detection model exported to ONNX, e.g. one trained to find tables, stamps or signatures:

from doctr.contrib import ArtefactDetector

detector = ArtefactDetector(model_path="path/to/your/model.onnx", labels=["table", "figure"])

Important

The model must be exported to ONNX with a dynamic batch size. Oriented bounding box (OBB) models are not supported.