xpark.dataset.ImageExtractTextLayout#

class xpark.dataset.ImageExtractTextLayout(mode: Literal['accurate', 'fast'] = 'accurate', text_confidence: float = 0.5, box_confidence: float = 0.6, concurrency: int = 10, use_cls: bool = False, text_detection_only: bool = False, engine: Literal['paddleocr', 'rapidocr'] = 'paddleocr', **ocr_kwargs)#

Extract the OCR text layout of an image via PP-OCRv6 (ONNXRuntime).

Each input row is one image; the output is a list of detected text regions TextLayoutBox. Input accepts a local path, a remote URI (cos:// / s3:// / http(s)://), raw image bytes, or a decoded uint8 HWC ndarray.

Parameters:

  • mode – "accurate" (default, best accuracy) or "fast" (faster, lower resource usage).

  • text_confidence – Recognition confidence threshold in [0, 1]; regions below it are dropped. Default 0.5.

  • box_confidence – Detection box confidence threshold in [0, 1]; boxes below it are dropped. Default 0.6 mirrors PaddleOCR’s text_det_box_thresh default.

  • concurrency – CPU inference threads (CPU only). Default 10.

  • use_cls – If True, run the text-line orientation classifier (rotates 180°-flipped lines upright) before recognition; keep False for upright text. Ignored when text_detection_only=True and engine="paddleocr", since the standalone detector has no orientation step. Default False.

  • text_detection_only – If True, only detect boxes and skip recognition entirely — for "paddleocr" a standalone detector is used (no recognition model is even loaded); for "rapidocr" the recognition step is disabled. text is "" and confidence is 0.0. Default False.

  • engine – OCR backend, "paddleocr" or "rapidocr". Default "paddleocr".

  • **ocr_kwargs – Extra params forwarded to the engine. "paddleocr" → PaddleOCR.__init__ (e.g. use_doc_orientation_classify / use_doc_unwarping); "rapidocr" → merged into the RapidOCR params dict (dotted keys, e.g. {"Det.box_thresh": 0.5}).

Examples

from xpark.dataset import ImageExtractTextLayout, from_items
from xpark.dataset.expressions import col

ds = from_items([{"image": "example.jpg"}])
ds = ds.with_column(
    "layout",
    ImageExtractTextLayout(text_confidence=0.75, box_confidence=0.6)
    .options(num_workers={"CPU": 1}, batch_size=8)
    .with_column(col("image")),
)

Methods

__call__(images)

Call self as a function.

options(**kwargs)

with_column(images)

__call__(images: pa.ChunkedArray) → pa.Array#

Call self as a function.

options(**kwargs: Unpack[ExprUDFOptions]) → Self#
with_column(images: pa.ChunkedArray) → pa.Array#