feat: font detection by comparing rendered vs OCR'd word widths

Co-Authored-By: Claude Opus 4.6 <[email protected]>

tests/test_font_detect.py

@@ -0,0 +1,38 @@
+from pathlib import Path
+
+from PIL import Image
+
+from unredact.pipeline.font_detect import detect_font, FontMatch
+from unredact.pipeline.ocr import ocr_page
+from unredact.pipeline.rasterize import rasterize_pdf
+
+
+def test_detect_font_returns_match(sample_pdf: Path):
+    pages = rasterize_pdf(sample_pdf, first_page=1, last_page=1)
+    lines = ocr_page(pages[0])
+    match = detect_font(lines, pages[0])
+    assert isinstance(match, FontMatch)
+    assert match.font_path is not None
+    assert match.font_size > 0
+    assert match.score > 0
+
+
+def test_detect_font_reasonable_size(sample_pdf: Path):
+    """At 300 DPI, body text in a letter-size doc is typically 30-60px."""
+    pages = rasterize_pdf(sample_pdf, first_page=1, last_page=1)
+    lines = ocr_page(pages[0])
+    match = detect_font(lines, pages[0])
+    # At 300 DPI, 11pt text ~ 46px. Allow wide range.
+    assert 20 < match.font_size < 80
+
+
+def test_font_match_can_render(sample_pdf: Path):
+    """The detected font should be usable for rendering."""
+    pages = rasterize_pdf(sample_pdf, first_page=1, last_page=1)
+    lines = ocr_page(pages[0])
+    match = detect_font(lines, pages[0])
+    font = match.to_pil_font()
+    # Should be able to measure text with it
+    bbox = font.getbbox("Hello")
+    assert bbox is not None
+    assert bbox[2] > 0  # width > 0

unredact/pipeline/font_detect.py

@@ -0,0 +1,132 @@
+from dataclasses import dataclass
+from pathlib import Path
+
+from PIL import Image, ImageDraw, ImageFont
+
+from unredact.pipeline.ocr import OcrLine
+
+# Candidate fonts to test (common document fonts available on the system)
+CANDIDATE_FONTS: list[str] = [
+    "Times New Roman",
+    "Arial",
+    "Courier New",
+    "Georgia",
+    "Liberation Serif",
+    "Liberation Sans",
+    "DejaVu Serif",
+    "DejaVu Sans",
+]
+
+# Size range to search (in pixels at rendering DPI)
+SIZE_RANGE = range(20, 80, 2)
+
+
+def _find_font_path(font_name: str) -> Path | None:
+    """Find the .ttf file for a font name using fc-match."""
+    import subprocess
+
+    result = subprocess.run(
+        ["fc-match", "--format=%{file}", font_name],
+        capture_output=True, text=True,
+    )
+    if result.returncode == 0 and result.stdout:
+        p = Path(result.stdout.strip())
+        if p.exists():
+            return p
+    return None
+
+
+@dataclass
+class FontMatch:
+    font_name: str
+    font_path: Path
+    font_size: int  # in pixels
+    score: float  # lower is better (mean absolute error in px)
+
+    def to_pil_font(self) -> ImageFont.FreeTypeFont:
+        return ImageFont.truetype(str(self.font_path), self.font_size)
+
+
+def _score_font(
+    font: ImageFont.FreeTypeFont,
+    lines: list[OcrLine],
+) -> float:
+    """Score how well a font matches the OCR'd character widths.
+
+    Returns mean absolute error in pixels between rendered and OCR'd word widths.
+    Lower is better.
+    """
+    errors: list[float] = []
+
+    for line in lines:
+        # Reconstruct words from chars
+        word = ""
+        word_start_x = -1
+        word_end_x = -1
+
+        for char in line.chars:
+            if char.text == " ":
+                if word and word_start_x >= 0:
+                    # Measure this word
+                    rendered_bbox = font.getbbox(word)
+                    rendered_w = rendered_bbox[2] - rendered_bbox[0]
+                    ocr_w = word_end_x - word_start_x
+                    if ocr_w > 0:
+                        errors.append(abs(rendered_w - ocr_w))
+                word = ""
+                word_start_x = -1
+            else:
+                if word_start_x < 0:
+                    word_start_x = char.x
+                word += char.text
+                word_end_x = char.x + char.w
+
+        # Last word in line
+        if word and word_start_x >= 0:
+            rendered_bbox = font.getbbox(word)
+            rendered_w = rendered_bbox[2] - rendered_bbox[0]
+            ocr_w = word_end_x - word_start_x
+            if ocr_w > 0:
+                errors.append(abs(rendered_w - ocr_w))
+
+    if not errors:
+        return float("inf")
+    return sum(errors) / len(errors)
+
+
+def detect_font(
+    lines: list[OcrLine],
+    page_image: Image.Image,
+) -> FontMatch:
+    """Detect the best matching font and size for OCR'd text.
+
+    Tries all candidate fonts at all sizes in SIZE_RANGE and returns
+    the combination with the lowest mean word-width error.
+    """
+    best: FontMatch | None = None
+
+    for font_name in CANDIDATE_FONTS:
+        font_path = _find_font_path(font_name)
+        if font_path is None:
+            continue
+
+        for size in SIZE_RANGE:
+            try:
+                font = ImageFont.truetype(str(font_path), size)
+            except Exception:
+                continue
+
+            score = _score_font(font, lines)
+
+            if best is None or score < best.score:
+                best = FontMatch(
+                    font_name=font_name,
+                    font_path=font_path,
+                    font_size=size,
+                    score=score,
+                )
+
+    if best is None:
+        raise RuntimeError("No matching font found. Check system fonts.")
+
+    return best