infer_document_orientation

This Ikomia plugin uses PaddleOCR's PP-LCNet_x1_0_doc_ori model to classify a document image into one of four clockwise orientation classes: 0, 90, 180, or 270 degrees.

Use with Ikomia API

Install Ikomia API

We recommend using a virtual environment.

pip install ikomia

Create a workflow

from ikomia.dataprocess.workflow import Workflow

wf = Workflow()
orientation = wf.add_task(
    name="infer_document_orientation",
    auto_connect=True,
)

wf.run_on(
    url=(
        "https://paddle-model-ecology.bj.bcebos.com/paddlex/"
        "imgs/demo_image/img_rot180_demo.jpg"
    )
)

display(orientation.get_image_with_graphics())
display(orientation.get_output(4).get_image())

Parameters

• model_name: PaddleOCR model name. Defaults to PP-LCNet_x1_0_doc_ori.
• cuda: use GPU 0 when enabled and the installed PaddlePaddle build supports CUDA. Defaults to False.
• enable_mkldnn: enable PaddlePaddle MKL-DNN CPU acceleration.
• cpu_threads: number of inference threads used on CPU. Defaults to 10.
• auto_rotate: rotate the second image output to an upright orientation using the predicted class. Defaults to False.

orientation.set_parameters(
    {
        "model_name": "PP-LCNet_x1_0_doc_ori",
        "cuda": "False",
        "enable_mkldnn": "True",
        "cpu_threads": "10",
        "auto_rotate": "True",
    }
)

Note: for GPU inference, replace the CPU paddlepaddle package with the appropriate paddlepaddle-gpu build for your CUDA environment, following PaddlePaddle's installation guide.

Output

Output 0 contains the predicted orientation label and its confidence score. The labels match PaddleOCR's four model categories:

Output 4 contains an image. When auto_rotate is enabled for whole-image classification, it is rotated to an upright orientation using PaddleX's counter-clockwise correction convention. Otherwise, it is an unchanged pass-through image.

References

• PaddleOCR document image orientation classification module
• PaddleOCR repository