Tesseract in practice

# Ubuntu / Debian
sudo apt install tesseract-ocr tesseract-ocr-hin tesseract-ocr-tam

# macOS
brew install tesseract tesseract-lang

# Windows: install the UB-Mannheim build, then add its folder to PATH

pip install pytesseract==0.3.13 pillow==11.0.0

import pytesseract
from PIL import Image

print("tesseract version:", pytesseract.get_tesseract_version())
print("languages installed:", pytesseract.get_languages())

page = Image.open("scan.png")

# --psm 3  = fully automatic page segmentation (the default)
# --oem 1  = LSTM engine only, which is what you want on Tesseract 4 and 5
text = pytesseract.image_to_string(page, lang="eng", config="--psm 3 --oem 1")
print(text)

# The same call, but asking for boxes and confidences instead of a blob of text.
tsv = pytesseract.image_to_data(page, lang="eng", config="--psm 3 --oem 1")
print(tsv)

# A tight crop of one line reads far better in single-line mode.
line = page.crop((36, 40, 356, 62))
print(pytesseract.image_to_string(line, config="--psm 7 --oem 1").strip())

# Restrict the alphabet when you know the field. This is the highest-value
# single flag on forms: an amount field cannot contain letters.
amount = page.crop((124, 96, 216, 114))
digits_only = "--psm 7 --oem 1 -c tessedit_char_whitelist=0123456789.,"
print(pytesseract.image_to_string(amount, config=digits_only).strip())

import csv
import io


def edit_distance(a, b):
    prev = list(range(len(b) + 1))
    for i, ca in enumerate(a, 1):
        cur = [i]
        for j, cb in enumerate(b, 1):
            cur.append(min(prev[j] + 1,           # deletion
                           cur[j - 1] + 1,        # insertion
                           prev[j - 1] + (ca != cb)))   # substitution
        prev = cur
    return prev[-1]


def cer(truth, guess):
    return edit_distance(truth, guess) / max(len(truth), 1)


def wer(truth, guess):
    t, g = truth.split(), guess.split()
    return edit_distance(t, g) / max(len(t), 1)


print("what different OCR mistakes actually cost you")
print(f"{'truth':<26}{'engine output':<26}{'CER':>7}{'WER':>7}")
cases = [
    ("INVOICE NO 4471", "INVOICE NO 4471", "perfect"),
    ("INVOICE NO 4471", "INVOICE NO 447l", "one digit-letter swap"),
    ("INVOICE NO 4471", "INVOICENO 4471", "one lost space"),
    ("INVOICE NO 4471", "1NVO1CE NO 4471", "two digit-letter swaps"),
    ("TOTAL 1250.00", "TOTAL 125O.OO", "zero read as capital O"),
    ("TOTAL 1250.00", "T0TAL 1250,00", "decimal point read as comma"),
]
for truth, guess, _ in cases:
    print(f"{truth:<26}{guess:<26}{cer(truth, guess):>7.3f}{wer(truth, guess):>7.3f}")

print("\nrow 2 costs 0.067 CER and 0.333 WER, for a single wrong character.")
print("row 6 costs 0.154 CER, and turns 1250.00 into something no parser accepts.")
print("CER is an average. It does not know which characters carry money.\n")

# Tesseract's image_to_data returns TSV in exactly this shape. These few rows are
# hand-written so the file runs without the Tesseract binary installed.
TSV = """level\tpage_num\tblock_num\tpar_num\tline_num\tword_num\tleft\ttop\twidth\theight\tconf\ttext
1\t1\t0\t0\t0\t0\t0\t0\t640\t480\t-1\t
2\t1\t1\t0\t0\t0\t36\t40\t320\t22\t-1\t
3\t1\t1\t1\t0\t0\t36\t40\t320\t22\t-1\t
4\t1\t1\t1\t1\t0\t36\t40\t320\t22\t-1\t
5\t1\t1\t1\t1\t1\t36\t40\t96\t18\t96\tINVOICE
5\t1\t1\t1\t1\t2\t140\t40\t34\t18\t95\tNO
5\t1\t1\t1\t1\t3\t182\t40\t62\t18\t31\t447l
5\t1\t1\t1\t1\t4\t250\t44\t8\t10\t12\t.
5\t1\t2\t1\t1\t1\t36\t96\t80\t18\t93\tTOTAL
5\t1\t2\t1\t1\t2\t124\t96\t92\t18\t88\t1250.00
"""

rows = list(csv.DictReader(io.StringIO(TSV), delimiter="\t", quoting=csv.QUOTE_NONE))
words = [r for r in rows if r["level"] == "5" and r["text"].strip()]
print(f"rows returned: {len(rows)}   of which words: {len(words)}")
print("the other rows are page, block, paragraph and line containers, with conf = -1\n")

for r in words:
    flag = "  <-- low confidence" if int(r["conf"]) < 60 else ""
    print(f"  conf {int(r['conf']):3d}   box ({r['left']},{r['top']},"
          f"{r['width']},{r['height']})   {r['text']!r}{flag}")

kept = [r for r in words if int(r["conf"]) >= 60]
print(f"\ndropping anything under 60 removes {len(words) - len(kept)} of {len(words)} words")

lines = {}
for r in kept:
    key = (r["block_num"], r["par_num"], r["line_num"])
    lines.setdefault(key, []).append(r)
print("\nrebuilt lines, sorted by block then line then word position:")
for key in sorted(lines):
    text = " ".join(w["text"] for w in sorted(lines[key], key=lambda w: int(w["left"])))
    print(f"  block {key[0]} line {key[2]}:  {text!r}")

truth = "INVOICE NO 4471 TOTAL 1250.00"
got = " ".join(" ".join(w["text"] for w in sorted(lines[k], key=lambda w: int(w["left"])))
                for k in sorted(lines))
print(f"\ntruth: {truth!r}")
print(f"kept:  {got!r}")
print(f"CER {cer(truth, got):.3f}   WER {wer(truth, got):.3f}")
print("filtering removed a stray '.', but it also removed the invoice number")