DocPreprocessor Pipeline

class DocPreprocessor:
    """
    Document preprocessing for orientation and distortion correction.

    Combines orientation classification and document unwarping
    to prepare scanned or photographed documents for OCR.
    """
    def __init__(
        self,
        doc_orientation_classify_model_name: str = None,
        doc_orientation_classify_model_dir: str = None,
        doc_unwarping_model_name: str = None,
        doc_unwarping_model_dir: str = None,
        use_doc_orientation_classify: bool = None,
        use_doc_unwarping: bool = None,
        paddlex_config: str = None,
        device: str = None,
        use_hpi: bool = None,
        **kwargs
    ):
        """
        Initialize DocPreprocessor pipeline.

        Args:
            doc_orientation_classify_model_name (str, optional): Orientation model
                Default: 'PP-LCNet_x1_0_doc_ori'
            doc_unwarping_model_name (str, optional): Unwarping model
                Default: 'UVDoc'
            use_doc_orientation_classify (bool, optional): Enable orientation correction
            use_doc_unwarping (bool, optional): Enable document unwarping
            device (str, optional): Device for inference ('cpu', 'gpu')
            use_hpi (bool, optional): Use high-performance inference
            paddlex_config (str or dict, optional): Configuration file or dict
        """

def predict(
    self,
    input,
    *,
    use_doc_orientation_classify: bool = None,
    use_doc_unwarping: bool = None
) -> list:
    """
    Preprocess document images.

    Args:
        input: Image/PDF path, numpy array, PIL Image, directory, or list
        use_doc_orientation_classify (bool, optional): Override orientation correction
        use_doc_unwarping (bool, optional): Override unwarping

    Returns:
        list: Preprocessed images with correction information
    """

def predict_iter(
    self,
    input,
    *,
    use_doc_orientation_classify: bool = None,
    use_doc_unwarping: bool = None
):
    """
    Iterate over preprocessing results for memory efficiency.

    Args:
        input: Image/PDF path, numpy array, PIL Image, directory, or list
        use_doc_orientation_classify (bool, optional): Override orientation correction
        use_doc_unwarping (bool, optional): Override unwarping

    Yields:
        dict: Preprocessed result for each image
    """

def close(self) -> None:
    """Close the pipeline and free resources."""

def export_paddlex_config_to_yaml(self, yaml_path: str) -> None:
    """Export configuration to YAML."""

from paddleocr import DocPreprocessor
import cv2

# Initialize with orientation correction only
preprocessor = DocPreprocessor(
    use_doc_orientation_classify=True,
    use_doc_unwarping=False
)

# Correct rotated documents
result = preprocessor.predict('rotated_doc.jpg')

# Save corrected image
corrected_img = result[0]['img']
cv2.imwrite('corrected_doc.jpg', corrected_img)

preprocessor.close()

from paddleocr import DocPreprocessor
import cv2

# Initialize with unwarping only
preprocessor = DocPreprocessor(
    use_doc_orientation_classify=False,
    use_doc_unwarping=True
)

# Unwarp curved or perspective-distorted document
result = preprocessor.predict('curved_doc.jpg')

# Save unwarped image
unwarped_img = result[0]['img']
cv2.imwrite('unwarped_doc.jpg', unwarped_img)

preprocessor.close()

from paddleocr import DocPreprocessor
import cv2

# Initialize with both corrections
preprocessor = DocPreprocessor(
    use_doc_orientation_classify=True,
    use_doc_unwarping=True
)

# Process document with both orientation and distortion issues
result = preprocessor.predict('phone_photo.jpg')

# Access corrected image
preprocessed = result[0]
print(f"Original orientation: {preprocessed.get('orientation_angle', 'N/A')}")
print(f"Applied unwarping: {preprocessed.get('unwarping_applied', False)}")

cv2.imwrite('preprocessed_doc.jpg', preprocessed['img'])

preprocessor.close()

from paddleocr import DocPreprocessor
import cv2
import os

preprocessor = DocPreprocessor(
    use_doc_orientation_classify=True,
    use_doc_unwarping=True
)

# Process directory of scanned documents
results = preprocessor.predict('scanned_docs/')

# Save all preprocessed images
os.makedirs('preprocessed/', exist_ok=True)
for result in results:
    input_name = os.path.basename(result['input_path'])
    output_path = os.path.join('preprocessed', input_name)
    cv2.imwrite(output_path, result['img'])

preprocessor.close()

from paddleocr import DocPreprocessor
import cv2

preprocessor = DocPreprocessor(
    use_doc_orientation_classify=True,
    use_doc_unwarping=True
)

# Process large batch with iterator
for result in preprocessor.predict_iter('large_document_set/'):
    # Process one at a time
    corrected = result['img']

    # Do something with the image
    cv2.imwrite(f"output/{result['input_path']}", corrected)

preprocessor.close()

from paddleocr import DocPreprocessor

preprocessor = DocPreprocessor()

# Only correct orientation for some images
rotated_results = preprocessor.predict(
    'rotated_docs/',
    use_doc_orientation_classify=True,
    use_doc_unwarping=False
)

# Only unwarp for others
curved_results = preprocessor.predict(
    'curved_docs/',
    use_doc_orientation_classify=False,
    use_doc_unwarping=True
)

preprocessor.close()

[
    {
        "input_path": "path/to/document.jpg",
        "img": numpy.ndarray,  # Preprocessed image (H, W, C)
        "orientation_angle": 90,  # Detected rotation (0, 90, 180, 270)
        "orientation_score": 0.98,  # Confidence of orientation detection
        "orientation_applied": True,  # Whether rotation was applied
        "unwarping_applied": True,  # Whether unwarping was applied
        "original_shape": (1024, 768, 3),  # Original image dimensions
        "preprocessed_shape": (1024, 768, 3)  # Final image dimensions
    }
]