📄 Awesome OCR multiple programing languages toolkits based on ONNX Runtime, OpenVINO, MNN, PaddlePaddle, TensorRT and PyTorch.
# -*- encoding: utf-8 -*-
import numpy as np
class ToMarkdown:
@classmethod
def to(cls, boxes, txts) -> str:
"""
根据 OCR 结果的坐标信息,将文本还原为近似原始排版的 Markdown。
Args:
result (RapidOCROutput): RapidOCR 的输出结果对象。
Returns:
str: 模拟原始排版的 Markdown 字符串。
"""
if boxes is None or txts is None:
return "没有检测到任何文本。"
items = [
{"text": text, "props": cls.get_box_properties(box)}
for box, text in zip(boxes, txts)
]
if not items:
return ""
items.sort(key=lambda item: (item["props"]["center_y"], item["props"]["left"]))
lines = []
for item in items:
matched_line = None
for line in lines:
if cls.is_same_line(item["props"], line["props"]):
matched_line = line
break
if matched_line is None:
lines.append({"items": [item], "props": dict(item["props"])})
continue
matched_line["items"].append(item)
matched_line["props"] = cls.merge_props(matched_line["props"], item["props"])
lines.sort(key=lambda line: (line["props"]["top"], line["props"]["left"]))
output_lines = []
prev_line_props = None
for line in lines:
line["items"].sort(key=lambda item: item["props"]["left"])
current_line_parts = [line["items"][0]["text"]]
prev_item_props = line["items"][0]["props"]
for item in line["items"][1:]:
current_props = item["props"]
gap = current_props["left"] - prev_item_props["right"]
current_line_parts.append(
cls.get_gap_text(gap, prev_item_props, current_props)
)
current_line_parts.append(item["text"])
prev_item_props = current_props
if prev_line_props is not None:
vertical_gap = line["props"]["top"] - prev_line_props["bottom"]
if (
vertical_gap
> max(prev_line_props["height"], line["props"]["height"]) * 0.8
):
output_lines.append("")
output_lines.append("".join(current_line_parts))
prev_line_props = line["props"]
return "\n".join(output_lines)
@staticmethod
def is_same_line(current_props: dict, line_props: dict) -> bool:
overlap_top = max(line_props["top"], current_props["top"])
overlap_bottom = min(line_props["bottom"], current_props["bottom"])
overlap = max(0.0, overlap_bottom - overlap_top)
min_height = max(1.0, min(current_props["height"], line_props["height"]))
overlap_ratio = overlap / min_height
center_diff = abs(current_props["center_y"] - line_props["center_y"])
return overlap_ratio > 0.5 or center_diff < min_height * 0.35
@staticmethod
def merge_props(a: dict, b: dict) -> dict:
top = min(a["top"], b["top"])
bottom = max(a["bottom"], b["bottom"])
left = min(a["left"], b["left"])
right = max(a["right"], b["right"])
return {
"top": top,
"bottom": bottom,
"left": left,
"right": right,
"height": bottom - top,
"width": right - left,
"center_y": top + (bottom - top) / 2,
}
@staticmethod
def get_gap_text(gap: float, prev_props: dict, current_props: dict) -> str:
if gap <= 1:
return ""
ref_width = max(1.0, min(prev_props["width"], current_props["width"]))
spaces = max(1, int(round(gap / max(1.0, ref_width * 0.5))))
return " " * min(spaces, 12)
@staticmethod
def get_box_properties(box: np.ndarray) -> dict:
"""从坐标数组中计算框的几何属性"""
# box shape is (4, 2) -> [[x1, y1], [x2, y2], [x3, y3], [x4, y4]]
ys = box[:, 1]
xs = box[:, 0]
top = np.min(ys)
bottom = np.max(ys)
left = np.min(xs)
right = np.max(xs)
return {
"top": top,
"bottom": bottom,
"left": left,
"right": right,
"height": bottom - top,
"width": right - left,
"center_y": top + (bottom - top) / 2,
}