paddleocr-api 0.0.1__py3-none-any.whl

paddleocr_api/models/optional_payload.py

@@ -0,0 +1,121 @@
+"""
+enum::AuxiliaryLayoutElement
+enum::LayoutShapeMode
+enum::PromptLabel
+json::OptionalPayload
+"""
+
+from numbers import Number
+from typing import List, TypedDict
+
+try:
+    from enum import StrEnum
+except ImportError:
+    from ..utils.enum import StrEnum
+
+
+class AuxiliaryLayoutElement(StrEnum):
+    """辅助版面元素"""
+    HEADER       = "header"       # 页眉
+    HEADER_IMAGE = "header_image" # 页眉图片
+    FOOTER       = "footer"       # 页脚
+    FOOTER_IMAGE = "footer_image" # 页脚图片
+    NUMBER       = "number"       # 页码
+    FOOTNOTE     = "footnote"     # 脚注
+    ASIDE_TEXT   = "aside_text"   # 旁注文本
+
+
+class LayoutShapeMode(StrEnum):
+    """版面检测结果的几何形状"""
+    AUTO = "auto" # 自动
+    RECT = "rect" # 矩形
+    QUAD = "quad" # 四边形
+    POLY = "poly" # 多边形
+
+
+class PromptLabel(StrEnum):
+    """prompt类型设置"""
+    OCR      = "ocr"      # 文本
+    FORMULA  = "formula"  # 公式
+    TABLE    = "table"    # 表格
+    CHART    = "chart"    # 图表
+    SEAL     = "seal"     # 印章
+    SPOTTING = "spotting" # 文本检测与识别
+
+
+class OptionalPayload(TypedDict):
+    """Optional parameters when creating jobs."""
+
+    # 辅助内容解析，默认为所有的 AuxiliaryLayoutElement 元素
+    # 模型会自动识别并过滤列出的辅助内容
+    markdownIgnoreLabels: List[AuxiliaryLayoutElement]
+
+    # 图片方向矫正, def. to `False`
+    # 开启后，可以自动识别并矫正 0°、90°、180°、270°的图片
+    useDocOrientationClassify: bool
+
+    # 图片扭曲矫正, def. to `False`
+    # 开启后，可以自动矫正扭曲图片，例如褶皱、倾斜等情况
+    useDocUnwarping: bool
+
+    # 版面分析, def. to `True`
+    # 开启后，系统会对文档进行版面分区和排序，并按区域类别识别。
+    # 如果您的文档仅包含文本、表格、公式或图表中的一种元素，可关闭此参数，直接进行对应识别
+    useLayoutDetection: bool
+
+    # 图表识别, def. to `False`
+    # 开启后，可以自动解析文档中的图表（如柱状图、饼图等）并转换为表格形式，方便查看和编辑数据
+    useChartRecognition: bool
+
+    # 印章识别, def. to `True`
+    # 开启后，可以识别文档中的印章内容，并将其提取为可编辑的文字
+    useSealRecognition: bool
+
+    # 图片文字识别, def. to `False`
+    # 开启后，对图片版面元素中的文本进行识别
+    useOcrForImageBlock: bool
+
+    # 跨页表格合并, def. to `True`
+    # 开启后，会识别跨页表格，将其合并为一个
+    mergeTables: bool
+
+    # 段落标题级别识别, def. to `True`
+    # 开启后，会识别段落标题级别
+    relevelTitles: bool
+
+    # 版面检测结果的几何形状, def. to `"auto"`
+    # 该参数决定了检测区域（如文本块、图片、表格等）边界的计算方式及展示形态。
+    # 默认为自动，系统根据检测目标的复杂程度和置信度，自动选择最合适的形状表达方式。
+    layoutShapeMode: LayoutShapeMode
+
+    # prompt类型设置, def. to `"ocr"`
+    # 当且仅当不使用版面分析模块时生效，设置输入文档的区域类型，当输入文档仅存在单个区域时推荐使用
+    promptLabel: PromptLabel
+
+    # 重复抑制强度, def. to `1.0`, recommended in [1.0, 1.2]
+    # 结果中出现重复文字、重复表格内容时，可适当调高
+    repetitionPenalty: Number
+
+    # 识别稳定性, def. to `0.0`, recommended in [0.0, 1.0]
+    # 结果不稳定或出现明显幻觉时调低，漏识别或者重复较多时可略微调高
+    temperature: Number
+
+    # 结果可信范围, def. to `1.0`
+    # 结果发散、不够可信时可适当调低，让模型更保守
+    topP: Number
+
+    # 图像最小总像素数
+    # 输入图片太小、文字看不清时可适当调高，一般无需调整
+    minPixels: int
+
+    # 图像最大总像素数
+    # 输入图片特别大、处理变慢或显存压力较大时可适当调低
+    maxPixels: int
+
+    # NMS后处理, def. to `True`
+    # 当且仅当使用版面分析模块时生效，开启后，会自动移除重复或高度重叠的区域框
+    layoutNms: bool
+
+    # 重构多页结果
+    # 对多页 pdf 解析结果进行重构，用于适配跨页表格合并和段落标题级别识别，默认初始化为False。
+    restructurePages: bool

paddleocr_api/models/result.py

@@ -0,0 +1,210 @@
+"""
+data::Markdown
+dataclass::Result
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional
+from typing_extensions import Self
+
+from ..utils.regex import MARKDOWN_TITLE_PATTERN
+
+
+@dataclass
+class PrunedResult:
+    """The recognition result of layout elements on a page."""
+    page_count: Optional[int]
+    width: int # page width
+    height: int # page height
+    model_settings: Dict[str, bool | List[str]] # Similar configuration in `OptionalPayload`.
+    parsing_res_list: List[Dict[str, Any]] # Layout recognition results
+    doc_preprocessor_res: Dict[str, int | Dict[str, bool]]
+    layout_det_res: Dict[str, List[Dict[str, Any]]]
+
+    @classmethod
+    def from_json(cls, data: Dict[str, Any]) -> PrunedResult:
+        """Construct from a JSON object."""
+        return PrunedResult(**data)
+
+
+@dataclass
+class Markdown:
+    """The recognition result of one page's markdown format."""
+    text: str
+    images: Dict[str, str] # image_path: image_url
+
+    @classmethod
+    def from_json(cls, data: Dict[str, str | Dict[str, str]]) -> Markdown:
+        """Construct from a JSON object."""
+        return Markdown(**data)
+
+
+@dataclass
+class LayoutParsingResult:
+    """One page layout recognition result"""
+    pruned_result: PrunedResult
+    markdown: Markdown
+    output_images: Dict[str, str] # output images when processing this page
+    input_image: str # The URL of the original image on this page
+
+    @classmethod
+    def from_json(cls, data: Dict[str, Any]) -> LayoutParsingResult:
+        """Construct from a JSON object."""
+        pruned_result = PrunedResult.from_json(data["prunedResult"])
+        markdown = Markdown.from_json(data["markdown"])
+        return LayoutParsingResult(
+            pruned_result=pruned_result,
+            markdown=markdown,
+            output_images=data["outputImages"],
+            input_image=data["inputImage"],
+        )
+
+
+@dataclass
+class PageSizeInfo:
+    """One page size information"""
+    width: int
+    height: int
+
+    @classmethod
+    def from_json(cls, data: Dict[str, int]) -> PageSizeInfo:
+        """Construct from a JSON object."""
+        return PageSizeInfo(**data)
+
+
+@dataclass
+class DataInfo:
+    """Processing result data"""
+    num_pages: int # Number of pages
+    pages: List[PageSizeInfo] # Size of each image
+    type: str # Input file format, like `"pdf"`
+
+    @classmethod
+    def from_json(cls, data: Dict[str, Any]) -> DataInfo:
+        """Construct from a JSON object."""
+        pages = list(map(PageSizeInfo.from_json, data["pages"]))
+        return DataInfo(
+            num_pages=data["numPages"],
+            pages=pages,
+            type=data["type"]
+        )
+
+
+
+_SENTENCE_END_SYMBOLS = ".!?。！？>"
+
+@dataclass
+class Result:
+    """Paddle OCR Processing Results for Documents"""
+    layout_parsing_results: List[LayoutParsingResult]
+    data_info: DataInfo
+    preprocessed_images: List[str] # Preprocessed Image URL List
+
+
+    def __add__(self, other: Result) -> Result:
+        """
+        Add the recognition results for the next section.
+        """
+        layout_parsing_results = self.layout_parsing_results + other.layout_parsing_results
+        preprocessed_images = self.preprocessed_images + other.preprocessed_images
+
+        data_info = self.data_info
+        other_data_info = other.data_info
+        num_pages = data_info.num_pages + other_data_info.num_pages
+        pages = data_info.pages + other_data_info.pages
+
+        return Result(
+            layout_parsing_results=layout_parsing_results,
+            preprocessed_images=preprocessed_images,
+            data_info=DataInfo(
+                num_pages=num_pages,
+                pages=pages,
+                type=data_info["type"],
+            )
+        )
+
+
+    @classmethod
+    def from_json(cls, data: Dict[str, Any]) -> Result:
+        """Construct from a JSON object."""
+        layout_parsing_results = list(map(LayoutParsingResult.from_json, data["layoutParsingResults"]))
+        data_info = DataInfo.from_json(data["dataInfo"])
+        return Result(
+            layout_parsing_results=layout_parsing_results,
+            data_info=data_info,
+            preprocessed_images=data["preprocessedImages"],
+        )
+
+
+    def extend(self, other: Result) -> Self:
+        """
+        Add the recognition results for the next section.
+        """
+        self.layout_parsing_results.extend(other.layout_parsing_results)
+        self.preprocessed_images.extend(other.preprocessed_images)
+
+        data_info = self.data_info
+        other_data_info = other.data_info
+        data_info.num_pages += other_data_info.num_pages
+        data_info.pages.extend(other_data_info.pages)
+
+        return self
+
+
+    @property
+    def markdown_text(self) -> str:
+        """
+        Return the parsing result of Markdown text format.
+
+        This will intelligently connect to the previous text.
+        """
+        # extract markdown
+        markdown_texts = []
+        for layout_parsing_result in self.layout_parsing_results:
+            markdown_text = layout_parsing_result.markdown.text
+            if len(markdown_texts) == 0:
+                markdown_texts.append(markdown_text)
+                continue
+
+            # Take out the last line of the previous text
+            last_row = markdown_texts[-1].rsplit("\n", 1)[-1].strip()
+            if last_row and (
+                MARKDOWN_TITLE_PATTERN.fullmatch(last_row) or # The previous line is the title
+                last_row.strip()[-1] in _SENTENCE_END_SYMBOLS # The previous line is a complete sentence
+            ):
+                markdown_texts.append(markdown_text)
+            else:
+                markdown_texts[-1] += ' ' + markdown_text
+
+        return "\n\n".join(markdown_texts)
+
+
+    @property
+    def markdown_images(self) -> Dict[str, str]:
+        """
+        Return the image in Markdown format in the parsed result.
+
+        Returns:
+            dict[str, str]:
+            - the key is image_relative_path.
+            - the value is image_url
+        """
+        # extract markdown
+        markdown_images = {}
+        for layout_parsing_result in self.layout_parsing_results:
+            markdown_images |= layout_parsing_result.markdown.images
+
+        return markdown_images
+
+
+    @property
+    def markdown(self) -> Markdown:
+        """
+        Return the Markdown formatted parsing result.
+        """
+        return Markdown(
+            text=self.markdown_text,
+            images=self.markdown_images
+        )

paddleocr_api/utils/enum.py

@@ -0,0 +1,8 @@
+"""
+type::StrEnum
+"""
+
+# 兼容 Python 3.10
+from enum import Enum
+class StrEnum(str, Enum):
+    """Enum where members are also (and must be) strings"""

paddleocr_api/utils/regex.py

@@ -0,0 +1,9 @@
+"""
+Regular expression patterns.
+"""
+
+import re
+
+URL_PATTERN = re.compile(r"https?:.*")
+
+MARKDOWN_TITLE_PATTERN = re.compile(r"\#+\s+.+")