fix

docs/pdf/check_legibility.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+"""Measure how legible each page of the photocopied manual is.
+
+    check_legibility.py copied.pdf [crisp.pdf]
+
+OCRs every page (tesseract) and scores it against the PDF's own invisible text
+layer, which is the known-correct text: the score is the share of true words
+that OCR recovers. If a crisp render is given, its score is the baseline (OCR
+is imperfect on monospace tables), and a page is judged by how much the copying
+cost it. Needs tesseract and poppler's pdftotext/pdftoppm. Development tool;
+the build does not run it.
+"""
+import difflib
+import os
+import re
+import subprocess
+import sys
+import tempfile
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+
+def words(text):
+    return re.findall(r"[A-Za-z0-9][A-Za-z0-9'.\-]*", text)
+
+
+def ocr_scores(pdf, dpi=130):
+    """Return {page: (recall, truth_word_count)} for each page of pdf."""
+    with tempfile.TemporaryDirectory() as tmp:
+        tmp = Path(tmp)
+        subprocess.run(["pdftoppm", "-r", str(dpi), "-gray", "-png", pdf, str(tmp / "p")], check=True)
+        images = sorted(tmp.glob("p-*.png"))
+
+        def score(img):
+            n = int(img.stem.split("-")[1])
+            truth = words(subprocess.run(["pdftotext", "-f", str(n), "-l", str(n), pdf, "-"],
+                                         capture_output=True, text=True, check=True).stdout)
+            got = words(subprocess.run(["tesseract", str(img), "-", "--psm", "6"],
+                                       capture_output=True, text=True, env={**os.environ, "OMP_THREAD_LIMIT": "1"}).stdout)
+            m = difflib.SequenceMatcher(None, truth, got, autojunk=False)
+            hit = sum(b.size for b in m.get_matching_blocks())
+            return n, (hit / max(1, len(truth)), len(truth))
+
+        with ThreadPoolExecutor(4) as pool:
+            return dict(pool.map(score, images))
+
+
+if __name__ == "__main__":
+    copied = ocr_scores(sys.argv[1])
+    crisp = ocr_scores(sys.argv[2]) if len(sys.argv) > 2 else {}
+    print("page  words  crisp  copied   loss")
+    losses = []
+    for n in sorted(copied):
+        c, w = copied[n]
+        b = crisp.get(n, (1.0, w))[0]
+        losses.append(b - c)
+        flag = "  <-- check" if b - c > 0.10 else ""
+        print(f"{n:4d} {w:6d}  {b:5.0%}  {c:5.0%}  {b - c:+5.0%}{flag}")
+    tot = sum(w for _, w in copied.values())
+    print(f"\nmean copied recall {sum(c * w for c, w in copied.values()) / tot:.1%}"
+          + (f", mean crisp {sum(c * w for c, w in crisp.values()) / tot:.1%}" if crisp else ""))