File size: 8,792 Bytes
1ef1de7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
"""Convert OCR lines into blocks, tables, figures, and reading order."""

from __future__ import annotations

import re

from .engines import EngineOutput
from .metrics import bbox_iou
from .normalization import normalize_ocr_text
from .schemas import Block, Document, Figure, Page, Table, TextLine, union_bbox
from .tables import infer_table_from_lines, looks_like_table_line


FIGURE_MARKER_RE = re.compile(r"^\s*(?:\[(?:figure|chart|image|चित्र)[^\]]*\]|(?:figure|fig\.?|chart|चित्र)\s*[::]?)", re.IGNORECASE)
CAPTION_RE = re.compile(r"^\s*(?:caption|figure|fig\.?|चित्र|तालिका)\s*[\s\d०-९IVXivx.-]*[::]", re.IGNORECASE)


def sort_lines_reading_order(lines: list[TextLine]) -> list[TextLine]:
    if not lines:
        return []
    heights = [line.bbox.h for line in lines if line.bbox.h > 0]
    line_height = sorted(heights)[len(heights) // 2] if heights else 20
    bucket_size = max(8, line_height * 0.75)
    return sorted(lines, key=lambda line: (round(line.bbox.y / bucket_size), line.bbox.x, line.bbox.y))


def build_document(source_path: str, output: EngineOutput) -> Document:
    pages: list[Page] = []
    tables = list(output.tables)
    figures = list(output.figures)
    for raw_page in output.pages:
        sorted_lines = sort_lines_reading_order(raw_page.text_lines)
        for index, line in enumerate(sorted_lines):
            if line.line_id is None:
                line.line_id = f"p{raw_page.page_index}-l{index}"
        blocks: list[Block] = []
        represented_figure_ids: set[str] = set()
        order = 0
        index = 0
        while index < len(sorted_lines):
            line = sorted_lines[index]
            if looks_like_figure_marker(line.text):
                marker_line = line
                caption_line: TextLine | None = None
                index += 1
                if index < len(sorted_lines) and _looks_like_caption_after_marker(marker_line, sorted_lines[index]):
                    caption_line = sorted_lines[index]
                    index += 1
                figure_lines = [marker_line] + ([caption_line] if caption_line is not None else [])
                figure_id = f"p{raw_page.page_index}-f{len(figures)}"
                caption = normalize_ocr_text(caption_line.text).strip() if caption_line is not None else None
                summary = normalize_ocr_text(marker_line.text).strip()
                figure = Figure(
                    figure_id=figure_id,
                    page_index=raw_page.page_index,
                    bbox=union_bbox([item.bbox for item in figure_lines]),
                    caption=caption,
                    summary=summary,
                    confidence=_average_confidence(figure_lines),
                    metadata={"line_ids": [item.line_id or "" for item in figure_lines]},
                )
                figures.append(figure)
                represented_figure_ids.add(figure_id)
                blocks.append(
                    Block(
                        block_id=f"p{raw_page.page_index}-b{order}",
                        block_type="figure",
                        page_index=raw_page.page_index,
                        bbox=figure.bbox,
                        order=order,
                        figure_id=figure_id,
                        confidence=figure.confidence,
                        line_ids=[item.line_id or "" for item in figure_lines],
                        text=caption or summary,
                    )
                )
                order += 1
                continue
            if looks_like_table_line(line.text):
                table_lines = [line]
                index += 1
                while index < len(sorted_lines) and looks_like_table_line(sorted_lines[index].text):
                    table_lines.append(sorted_lines[index])
                    index += 1
                table_id = f"p{raw_page.page_index}-t{len(tables)}"
                table = infer_table_from_lines(table_lines, table_id)
                existing_table = _find_overlapping_table(tables, table)
                represented_table = existing_table or table
                if existing_table is None:
                    tables.append(table)
                blocks.append(
                    Block(
                        block_id=f"p{raw_page.page_index}-b{order}",
                        block_type="table",
                        page_index=raw_page.page_index,
                        bbox=represented_table.bbox,
                        order=order,
                        table_id=represented_table.table_id,
                        confidence=represented_table.confidence,
                        line_ids=[line.line_id or "" for line in table_lines],
                    )
                )
                order += 1
                continue

            text_lines = [line]
            index += 1
            while (
                index < len(sorted_lines)
                and not looks_like_table_line(sorted_lines[index].text)
                and not looks_like_figure_marker(sorted_lines[index].text)
            ):
                previous = text_lines[-1]
                current = sorted_lines[index]
                vertical_gap = current.bbox.y - previous.bbox.bottom
                if vertical_gap > max(previous.bbox.h, 24) * 1.5:
                    break
                text_lines.append(current)
                index += 1
            text = "\n".join(normalize_ocr_text(item.text).strip() for item in text_lines if item.text.strip())
            if not text.strip():
                continue
            confidence = _average_confidence(text_lines)
            block_type = "title" if order == 0 and len(text) <= 120 and len(text_lines) == 1 else "text"
            blocks.append(
                Block(
                    block_id=f"p{raw_page.page_index}-b{order}",
                    block_type=block_type,
                    page_index=raw_page.page_index,
                    bbox=union_bbox([item.bbox for item in text_lines]),
                    order=order,
                    text=text,
                    confidence=confidence,
                    line_ids=[item.line_id or "" for item in text_lines],
                )
            )
            order += 1

        for figure in [item for item in figures if item.page_index == raw_page.page_index and item.figure_id not in represented_figure_ids]:
            blocks.append(
                Block(
                    block_id=f"p{raw_page.page_index}-b{order}",
                    block_type="figure",
                    page_index=raw_page.page_index,
                    bbox=figure.bbox,
                    order=order,
                    figure_id=figure.figure_id,
                    confidence=figure.confidence,
                    text=figure.caption or figure.summary or "",
                )
            )
            order += 1
        blocks = sorted(blocks, key=lambda block: (block.bbox.y, block.bbox.x, block.order))
        for order_index, block in enumerate(blocks):
            block.order = order_index
        pages.append(
            Page(
                page_index=raw_page.page_index,
                width=raw_page.width,
                height=raw_page.height,
                text_lines=sorted_lines,
                blocks=blocks,
                metadata=raw_page.metadata,
            )
        )
    return Document(source_path=source_path, pages=pages, tables=tables, figures=figures, metadata=dict(output.metadata))


def looks_like_figure_marker(text: str) -> bool:
    return bool(FIGURE_MARKER_RE.match(text.strip()))


def looks_like_caption(text: str) -> bool:
    return bool(CAPTION_RE.match(text.strip()))


def _looks_like_caption_after_marker(marker_line: TextLine, candidate: TextLine) -> bool:
    if looks_like_caption(candidate.text):
        return True
    if looks_like_figure_marker(candidate.text) or looks_like_table_line(candidate.text):
        return False
    if not candidate.text.strip():
        return False
    vertical_gap = candidate.bbox.y - marker_line.bbox.bottom
    max_gap = max(marker_line.bbox.h, candidate.bbox.h, 24) * 1.5
    return 0 <= vertical_gap <= max_gap


def _find_overlapping_table(tables: list[Table], candidate: Table, *, iou_threshold: float = 0.9) -> Table | None:
    for table in tables:
        if table.page_index != candidate.page_index:
            continue
        if bbox_iou(table.bbox, candidate.bbox) >= iou_threshold:
            return table
    return None


def _average_confidence(lines: list[TextLine]) -> float | None:
    confidences = [item.confidence for item in lines if item.confidence is not None]
    return sum(confidences) / len(confidences) if confidences else None