File size: 6,855 Bytes
4837bb3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""OCR evaluation for outputs produced by controlnet_benchmark.py."""
from __future__ import annotations

import argparse
import json
import math
from pathlib import Path

from PIL import Image, ImageChops, ImageDraw, ImageFilter, ImageFont, ImageOps

from benchmarks.text_accuracy import (
    centered_text_crop,
    character_error_rate,
    normalize_ocr,
    run_ocr,
)


def _normalized_outline(
    image: Image.Image,
    size: tuple[int, int] = (384, 128),
    *,
    light_ink: bool = False,
) -> Image.Image:
    grayscale = image.convert("L")
    if light_ink:
        ink = grayscale.point(lambda value: 255 if value >= 128 else 0)
    else:
        ink = grayscale.point(lambda value: 255 if value < 128 else 0)
    bbox = ink.getbbox()
    if bbox is None:
        return Image.new("L", size, 0)
    cropped = ImageOps.expand(ink.crop(bbox), border=4, fill=0)
    normalized = cropped.resize(size, Image.Resampling.LANCZOS)
    normalized = normalized.point(lambda value: 255 if value >= 128 else 0)
    outer = normalized.filter(ImageFilter.MaxFilter(3))
    inner = normalized.filter(ImageFilter.MinFilter(3))
    return ImageChops.subtract(outer, inner)


def glyph_similarity(image: Image.Image, control: Image.Image) -> float:
    generated = _normalized_outline(image.crop(centered_text_crop(image)))
    expected = _normalized_outline(control, light_ink=True)
    generated_dilated = generated.filter(ImageFilter.MaxFilter(7))
    expected_dilated = expected.filter(ImageFilter.MaxFilter(7))
    generated_pixels = sum(1 for value in generated.getdata() if value)
    expected_pixels = sum(1 for value in expected.getdata() if value)
    if not generated_pixels or not expected_pixels:
        return 0.0
    generated_hit = sum(
        1
        for edge, nearby in zip(generated.getdata(), expected_dilated.getdata())
        if edge and nearby
    )
    expected_hit = sum(
        1
        for edge, nearby in zip(expected.getdata(), generated_dilated.getdata())
        if edge and nearby
    )
    precision = generated_hit / generated_pixels
    recall = expected_hit / expected_pixels
    return round(2 * precision * recall / max(1e-9, precision + recall), 4)


def light_ink_margin(image: Image.Image) -> int:
    ink = image.convert("L").point(lambda value: 255 if value >= 128 else 0)
    bbox = ink.getbbox()
    if bbox is None:
        return -1
    width, height = image.size
    return min(bbox[0], bbox[1], width - bbox[2], height - bbox[3])


def write_contact_sheets(
    results: list[dict[str, object]],
    output_dir: Path,
    *,
    columns: int = 5,
    rows: int = 4,
) -> list[str]:
    page_size = columns * rows
    thumb_size = 240
    label_height = 44
    font_path = Path("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf")
    font = ImageFont.truetype(str(font_path), 22) if font_path.exists() else ImageFont.load_default()
    names: list[str] = []
    for page_index in range(math.ceil(len(results) / page_size)):
        page_rows = results[page_index * page_size : (page_index + 1) * page_size]
        sheet = Image.new(
            "RGB",
            (columns * thumb_size, rows * (thumb_size + label_height)),
            "white",
        )
        draw = ImageDraw.Draw(sheet)
        for cell_index, result in enumerate(page_rows):
            image_path = output_dir / str(result["image"])
            with Image.open(image_path) as source:
                thumbnail = source.convert("RGB")
                thumbnail.thumbnail((thumb_size, thumb_size), Image.Resampling.LANCZOS)
            x = (cell_index % columns) * thumb_size
            y = (cell_index // columns) * (thumb_size + label_height)
            sheet.paste(
                thumbnail,
                (x + (thumb_size - thumbnail.width) // 2, y),
            )
            draw.text((x + 6, y + thumb_size + 4), str(result["text"]), fill="black", font=font)
        name = f"contact-sheet-{page_index + 1:02d}.jpg"
        sheet.save(output_dir / name, quality=90)
        names.append(name)
    return names


def summarize(results: list[dict[str, object]]) -> dict[str, dict[str, float | int]]:
    strategies = sorted({str(result["strategy"]) for result in results})
    summary: dict[str, dict[str, float | int]] = {}
    for strategy in strategies:
        rows = [result for result in results if result["strategy"] == strategy]
        summary[strategy] = {
            "samples": len(rows),
            "exact": sum(bool(row["exact_match"]) for row in rows),
            "mean_cer": round(sum(float(row["cer"]) for row in rows) / len(rows), 4),
            "mean_glyph_similarity": round(
                sum(float(row.get("glyph_similarity", 0.0)) for row in rows) / len(rows),
                4,
            ),
            "edge_touching_controls": sum(
                int(row.get("control_margin_px", 1)) <= 0 for row in rows
            ),
        }
    return summary


def run(report_path: Path) -> dict[str, object]:
    generation = json.loads(report_path.read_text(encoding="utf-8"))
    output_dir = report_path.parent
    results: list[dict[str, object]] = []
    for row in generation["results"]:
        image_path = output_dir / row["image"]
        with Image.open(image_path) as image:
            crop = centered_text_crop(image)
        actual = run_ocr(image_path, "rus+eng", crop=crop)
        gold = normalize_ocr(row["text"])
        similarity = 0.0
        control_margin_px: int | None = None
        if row.get("control"):
            with Image.open(image_path) as image, Image.open(output_dir / row["control"]) as control:
                similarity = glyph_similarity(image, control)
                control_margin_px = light_ink_margin(control)
        results.append(
            {
                **row,
                "gold": gold,
                "ocr": actual,
                "exact_match": actual == gold,
                "cer": round(character_error_rate(gold, actual), 4),
                "glyph_similarity": similarity,
                "control_margin_px": control_margin_px,
            }
        )
    contact_sheets = write_contact_sheets(results, output_dir)
    report: dict[str, object] = {
        "generation_report": report_path.name,
        "summary": summarize(results),
        "contact_sheets": contact_sheets,
        "results": results,
    }
    (output_dir / "accuracy-report.json").write_text(
        json.dumps(report, ensure_ascii=False, indent=2),
        encoding="utf-8",
    )
    return report


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("generation_report", type=Path)
    return parser.parse_args()


if __name__ == "__main__":
    args = parse_args()
    completed = run(args.generation_report.resolve())
    print(json.dumps(completed, ensure_ascii=False))