rerun-io--rerun/examples/python/ocr/ocr.py

#!/usr/bin/env python3
"""OCR template."""

from __future__ import annotations

import argparse
import logging
import os
from enum import Enum
from pathlib import Path
from typing import TYPE_CHECKING, Any, Final, TypeAlias

import cv2
import numpy as np
import numpy.typing as npt
import pandas as pd
import pdf2image
import requests
import tqdm
from paddleocr import PPStructure
from paddleocr.ppstructure.recovery.recovery_to_doc import sorted_layout_boxes

import rerun as rr  # pip install rerun-sdk
import rerun.blueprint as rrb

if TYPE_CHECKING:
    from collections.abc import Iterable

EXAMPLE_DIR: Final = Path(os.path.dirname(__file__))
DATASET_DIR: Final = EXAMPLE_DIR / "dataset"

SAMPLE_IMAGE_URLs = ["https://storage.googleapis.com/rerun-example-datasets/ocr/paper.png"]

LayoutStructure: TypeAlias = tuple[
    list[str],
    list[str],
    list[rrb.Spatial2DView],
    list[rrb.Spatial2DView],
    list[rrb.Spatial2DView],
]

# Supportive Classes


class Color:
    Red = (255, 0, 0)
    Green = (0, 255, 0)
    Blue = (0, 0, 255)
    Yellow = (255, 255, 0)
    Cyan = (0, 255, 255)
    Magenta = (255, 0, 255)
    Purple = (128, 0, 128)
    Orange = (255, 165, 0)


"""
LayoutType:
    Defines an enumeration for different types of document layout elements, each associated with a unique number, name,
    and color. Types:
    - UNKNOWN: Default type for undefined or unrecognized elements, represented by purple.
    - TITLE: Represents the title of a document, represented by red.
    - TEXT: Represents plain text content within the document, represented by green.
    - FIGURE: Represents graphical or image content, represented by blue.
    - FIGURE_CAPTION: Represents captions for figures, represented by yellow.
    - TABLE: Represents tabular data, represented by cyan.
    - TABLE_CAPTION: Represents captions for tables, represented by magenta.
    - REFERENCE: Represents citation references within the document, also represented by purple.
    - Footer: Represents footer of the document, represented as orange.
"""


class LayoutType(Enum):
    UNKNOWN = (0, "unknown", Color.Purple)
    TITLE = (1, "title", Color.Red)
    TEXT = (2, "text", Color.Green)
    FIGURE = (3, "figure", Color.Blue)
    FIGURE_CAPTION = (4, "figure_caption", Color.Yellow)
    TABLE = (5, "table", Color.Cyan)
    TABLE_CAPTION = (6, "table_caption", Color.Magenta)
    REFERENCE = (7, "reference", Color.Purple)
    FOOTER = (8, "footer", Color.Orange)

    def __str__(self) -> str:
        return str(self.value[1])  # Returns the string part (type)

    @property
    def number(self) -> int:
        return self.value[0]  # Returns the numerical identifier

    @property
    def type(self) -> str:
        return self.value[1]  # Returns the type

    @property
    def color(self) -> tuple[int, int, int]:
        return self.value[2]  # Returns the color

    @staticmethod
    def get_class_id(text: str) -> int:
        try:
            return LayoutType[text.upper()].number
        except KeyError:
            logging.warning(f"Invalid layout type {text}")
            return 0

    @staticmethod
    def get_type(text: str) -> LayoutType:
        try:
            return LayoutType[text.upper()]
        except KeyError:
            logging.warning(f"Invalid layout type {text}")
            return LayoutType.UNKNOWN

    @classmethod
    def get_annotation(cls) -> list[tuple[int, str, tuple[int, int, int]]]:
        return [(layout.number, layout.type, layout.color) for layout in cls]


"""
Layout Class:
    The main purpose of this class is to:
    1. Keep track of the layout types (including type, numbering)
    2. Save the detections for each layout (text, img or table)
    3. Save the bounding box of each detected layout
    4. Generate the recovery text document
"""


class Layout:
    def __init__(self, page_number: int, show_unknown: bool = False) -> None:
        self.counts = dict.fromkeys(LayoutType, 0)
        self.records: dict[LayoutType, Any] = {layout_type: [] for layout_type in LayoutType}
        self.recovery = """"""
        self.page_number = page_number
        self.show_unknown = show_unknown

    def add(
        self,
        layout_type: LayoutType,
        bounding_box: list[int],
        detections: Iterable[dict[str, Any]] | None = None,
        table: str | None = None,
        img: dict[str, Any] | None = None,
    ) -> None:
        if layout_type in LayoutType:
            self.counts[layout_type] += 1
            name = f"{layout_type}{self.counts[layout_type]}"
            logging.info(f"Saved layout type {layout_type} with name: {name}")
            self.records[layout_type].append({
                "type": layout_type,
                "name": name,
                "bounding_box": bounding_box,
                "detections": detections,
                "table": table,
            })
            if layout_type != LayoutType.UNKNOWN or self.show_unknown:  # Discards the unknown layout types detections
                path = f"recording://page_{self.page_number}/Image/{layout_type.type.title()}/{name.title()}"
                self.recovery += f"\n\n## [{name.title()}]({path})\n\n"  # Log Type as Heading
                if layout_type == LayoutType.TABLE:
                    if table:
                        self.recovery += table  # Log details (table)
                elif detections:
                    for index, detection in enumerate(detections):
                        path_text = f"recording://page_{self.page_number}/Image/{layout_type.type.title()}/{name.title()}/Detections/{index}"
                        self.recovery += f" [{detection['text']}]({path_text})"  # Log details (text)
        else:
            logging.warning(f"Invalid layout type detected: {layout_type}")

    def get_count(self, layout_type: LayoutType) -> int:
        if layout_type in LayoutType:
            return self.counts[layout_type]
        else:
            raise ValueError("Invalid layout type")

    def get_records(self) -> dict[LayoutType, list[dict[str, Any]]]:
        return self.records

    def save_all_layouts(self, results: list[dict[str, Any]]) -> None:
        for line in results:
            self.save_layout_data(line)
        for layout_type in LayoutType:
            logging.info(f"Number of detections for type {layout_type}: {self.counts[layout_type]}")

    def save_layout_data(self, line: dict[str, Any]) -> None:
        type = line.get("type", "empty")
        box = line.get("bbox", [0, 0, 0, 0])
        layout_type = LayoutType.get_type(type)
        detections, table, img = [], None, None
        if layout_type == LayoutType.TABLE:
            table = self.get_table_markdown(line)
        elif layout_type == LayoutType.FIGURE:
            detections = self.get_detections(line)
            img = line.get("img")  # Currently not in use
        else:
            detections = self.get_detections(line)
        self.add(layout_type, box, detections=detections, table=table, img=img)

    @staticmethod
    def get_detections(line: dict[str, Any]) -> list[dict[str, Any]]:
        detections = []
        results = line.get("res")
        if results is not None:
            for i, result in enumerate(results):
                text = result.get("text")
                confidence = result.get("confidence")
                box = result.get("text_region")
                x_min, y_min = box[0]
                x_max, y_max = box[2]
                new_box = [x_min, y_min, x_max, y_max]
                detections.append({"id": i, "text": text, "confidence": confidence, "box": new_box})
        return detections

    # Safely attempt to extract the HTML table from the results
    @staticmethod
    def get_table_markdown(line: dict[str, Any]) -> str:
        try:
            html_table = line.get("res", {}).get("html")
            if not html_table:
                return "No table found."

            dataframes = pd.read_html(html_table)
            if not dataframes:
                return "No data extracted from the table."

            markdown_table = dataframes[0].to_markdown()
            return markdown_table  # type: ignore[no-any-return]

        except Exception as e:
            return f"Error processing the table: {e!s}"


def process_layout_records(layout: Layout, page_path: str) -> LayoutStructure:
    paths, detections_paths = [], []
    zoom_paths: list[rrb.Spatial2DView] = []
    zoom_paths_figures: list[rrb.Spatial2DView] = []
    zoom_paths_tables: list[rrb.Spatial2DView] = []
    zoom_paths_texts: list[rrb.Spatial2DView] = []

    for layout_type in LayoutType:
        for record in layout.records[layout_type]:
            record_name = record["name"].title()
            record_base_path = f"{page_path}/Image/{layout_type.type.title()}/{record_name}"
            paths.append(f"-{record_base_path}/**")
            detections_paths.append(f"-{record_base_path}/Detections/**")

            # Log bounding box
            rr.log(
                record_base_path,
                rr.Boxes2D(
                    array=record["bounding_box"],
                    array_format=rr.Box2DFormat.XYXY,
                    labels=[str(layout_type.type)],
                    class_ids=[str(layout_type.number)],
                ),
                rr.AnyValues(name=record_name),
            )

            log_detections(layout_type, record, record_base_path)

            # Prepare zoom path views
            update_zoom_paths(
                layout,
                layout_type,
                record,
                paths,
                page_path,
                zoom_paths,
                zoom_paths_figures,
                zoom_paths_tables,
                zoom_paths_texts,
            )

    return paths, detections_paths, zoom_paths_figures, zoom_paths_tables, zoom_paths_texts


def log_detections(layout_type: LayoutType, record: dict[str, Any], page_path: str) -> None:
    if layout_type == LayoutType.TABLE:
        rr.log(f"Extracted{record['name']}", rr.TextDocument(record["table"], media_type=rr.MediaType.MARKDOWN))
    else:
        for detection in record.get("detections", []):
            rr.log(
                f"{page_path}/Detections/{detection['id']}",
                rr.Boxes2D(
                    array=detection["box"],
                    array_format=rr.Box2DFormat.XYXY,
                    class_ids=[str(layout_type.number)],
                ),
                rr.AnyValues(DetectionID=detection["id"], Text=detection["text"], Confidence=detection["confidence"]),
            )


def update_zoom_paths(
    layout: Layout,
    layout_type: LayoutType,
    record: dict[str, Any],
    paths: list[str],
    page_path: str,
    zoom_paths: list[rrb.Spatial2DView],
    zoom_paths_figures: list[rrb.Spatial2DView],
    zoom_paths_tables: list[rrb.Spatial2DView],
    zoom_paths_texts: list[rrb.Spatial2DView],
) -> None:
    if layout_type in [LayoutType.FIGURE, LayoutType.TABLE, LayoutType.TEXT]:
        current_paths = paths.copy()
        current_paths.remove(f"-{page_path}/Image/{layout_type.type.title()}/{record['name'].title()}/**")
        bounds = rrb.VisualBounds2D(
            x_range=[record["bounding_box"][0] - 10, record["bounding_box"][2] + 10],
            y_range=[record["bounding_box"][1] - 10, record["bounding_box"][3] + 10],
        )

        # Add to zoom paths
        view = rrb.Spatial2DView(
            name=record["name"].title(),
            contents=[f"{page_path}/Image/**", *current_paths],
            visual_bounds=bounds,
        )
        zoom_paths.append(view)

        # Add to type-specific zoom paths
        if layout_type == LayoutType.FIGURE:
            zoom_paths_figures.append(view)
        elif layout_type == LayoutType.TABLE:
            zoom_paths_tables.append(view)
        elif layout_type != LayoutType.UNKNOWN or layout.show_unknown:
            zoom_paths_texts.append(view)


def generate_blueprint(
    layouts: list[Layout],
    processed_layouts: list[LayoutStructure],
) -> rrb.Blueprint:
    page_tabs = []
    for layout, processed_layout in zip(layouts, processed_layouts, strict=False):
        paths, detections_paths, zoom_paths_figures, zoom_paths_tables, zoom_paths_texts = processed_layout

        section_tabs = []
        content_data: dict[str, Any] = {
            "Figures": zoom_paths_figures,
            "Tables": zoom_paths_tables,
            "Texts": zoom_paths_texts,
        }

        for name, paths in content_data.items():
            if paths:
                section_tabs.append(rrb.Tabs(*paths, name=name))  # type: ignore[arg-type]

        page_path = f"page_{layout.page_number}"
        page_tabs.append(
            rrb.Vertical(
                rrb.Horizontal(
                    rrb.Spatial2DView(
                        name="Layout",
                        origin=f"{page_path}/Image/",
                        contents=[f"{page_path}/Image/**", *detections_paths],
                    ),
                    rrb.Spatial2DView(name="Detections", contents=[f"{page_path}/Image/**"]),
                    rrb.TextDocumentView(name="Recovery", contents=f"{page_path}/Recovery"),
                ),
                rrb.Horizontal(*section_tabs),
                name=page_path,
                row_shares=[4, 3],
            ),
        )

    return rrb.Blueprint(
        rrb.Tabs(*page_tabs),
        collapse_panels=True,
    )


def detect_and_log_layouts(file_path: str) -> None:
    images: list[npt.NDArray[np.uint8]] = []
    if file_path.endswith(".pdf"):
        # convert pdf to images
        images.extend(np.array(img, dtype=np.uint8) for img in pdf2image.convert_from_path(file_path))
    else:
        # read image
        img = cv2.imread(file_path)
        image_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)  # Rerun can handle BGR as well, but `ocr_model_pp` expects RGB
        images.append(image_rgb.astype(np.uint8))

    # Extracte the layout from each image
    layouts: list[Layout] = []
    page_numbers = [i + 1 for i in range(len(images))]
    processed_layouts: list[LayoutStructure] = []
    for image, page_number in zip(images, page_numbers, strict=False):
        layouts.append(detect_and_log_layout(image, page_number))
        page_path = f"page_{page_number}"

        # Generate and send a blueprint based on the detected layouts
        processed_layouts.append(
            process_layout_records(
                layouts[-1],
                page_path,
            ),
        )
        logging.info("Sending blueprint…")
        blueprint = generate_blueprint(layouts, processed_layouts)
        rr.send_blueprint(blueprint)
        logging.info("Blueprint sent…")


def detect_and_log_layout(image_rgb: npt.NDArray[np.uint8], page_number: int) -> Layout:
    # Layout Object - This will contain the detected layouts and their detections
    layout = Layout(page_number)
    page_path = f"page_{page_number}"

    # Log Image and add Annotation Context
    rr.log(f"{page_path}/Image", rr.Image(image_rgb))
    rr.log(
        f"{page_path}/Image",
        # The annotation is defined in the Layout class based on its properties
        rr.AnnotationContext(LayoutType.get_annotation()),
        static=True,
    )

    # Paddle Model - Getting Predictions
    logging.info("Start detection… (It usually takes more than 10-20 seconds per page)")
    ocr_model_pp = PPStructure(show_log=False, recovery=True)
    logging.info("model loaded")
    result_pp = ocr_model_pp(image_rgb)
    _, w, _ = image_rgb.shape
    result_pp = sorted_layout_boxes(result_pp, w)
    logging.info("Detection finished…")

    # Add results to the layout
    layout.save_all_layouts(result_pp)
    logging.info("All results are saved…")

    # Recovery Text Document for the detected text
    rr.log(f"{page_path}/Recovery", rr.TextDocument(layout.recovery, media_type=rr.MediaType.MARKDOWN))

    return layout


def get_downloaded_path(dataset_dir: Path, video_name: str) -> str:
    video_file_name = f"{video_name}.mp4"
    destination_path = dataset_dir / video_file_name
    if destination_path.exists():
        logging.info("%s already exists. No need to download", destination_path)
        return str(destination_path)
    else:
        logging.warning("Problem on image downloading")
        return ""


def download_file(url: str, path: Path) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    logging.info("Downloading %s to %s", url, path)
    response = requests.get(url, stream=True)
    with tqdm.tqdm.wrapattr(
        open(path, "wb"),
        "write",
        miniters=1,
        total=int(response.headers.get("content-length", 0)),
        desc=f"Downloading {path.name}",
    ) as f:
        for chunk in response.iter_content(chunk_size=4096):
            f.write(chunk)


def main() -> None:
    parser = argparse.ArgumentParser(
        description="OCR Example - Layout Analysis and Text Detections. It automatically downloads the PaddleOCR libraries and models.",
    )

    parser.add_argument(
        "--demo-image",
        type=str,
        default="paper",
        choices=["paper"],
        help="Run on a demo image automatically downloaded",
    )
    parser.add_argument(
        "--file",
        type=str,
        help="Run on the provided image/pdf (for pdf files `poppler` must be installed)",
    )

    rr.script_add_args(parser)
    args = parser.parse_args()

    rr.script_setup(
        args,
        "rerun_ocr_example",
        default_blueprint=rrb.Blueprint(
            rrb.Vertical(
                rrb.Spatial2DView(name="Input", contents=["Image/**"]),
            ),
            collapse_panels=True,
        ),
    )
    rr.script_teardown(args)

    logging.getLogger().addHandler(rr.LoggingHandler("logs/handler"))
    logging.getLogger().setLevel(-1)

    # Choose the appropriate run mode based on provided arguments
    if args.file:
        detect_and_log_layouts(args.file)
    else:
        img_path = DATASET_DIR / f"{args.demo_image}.png"
        if not img_path.exists():
            download_file(SAMPLE_IMAGE_URLs[0], img_path)
        detect_and_log_layouts(str(img_path))


if __name__ == "__main__":
    main()