OCR 6: đo sai số ký tự thay vì chỉ nhìn kết quả

from pathlib import Path
import csv
import shutil
import unicodedata
import cv2
import numpy as np
import matplotlib.pyplot as plt
import pytesseract

output = Path("ocr_error_output")
output.mkdir(exist_ok=True)

def normalize(text):
    return unicodedata.normalize("NFC", text).strip()

def edit_distance(a, b):
    previous = list(range(len(b)+1))
    for i, char_a in enumerate(a, start=1):
        current = [i]
        for j, char_b in enumerate(b, start=1):
            current.append(min(current[-1]+1, previous[j]+1,
                               previous[j-1]+(char_a != char_b)))
        previous = current
    return previous[-1]

def cer(reference, prediction):
    reference, prediction = normalize(reference), normalize(prediction)
    if not reference:
        raise ValueError("CER can nhan dung khong rong")
    return edit_distance(reference, prediction) / len(reference)

assert edit_distance("ABC", "ADC") == 1
assert edit_distance("ABC", "ABXC") == 1
assert edit_distance("ABC", "AC") == 1
assert cer("ABC", "ABC") == 0
assert cer("A", "AAAA") == 3

engine = shutil.which("tesseract")
if not engine:
    candidate = Path(r"C:\Program Files\Tesseract-OCR\tesseract.exe")
    engine = str(candidate) if candidate.exists() else None
if not engine:
    raise RuntimeError("Can cai Tesseract OCR va du lieu eng, sau do kiem tra duong dan")
pytesseract.pytesseract.tesseract_cmd = engine
reference = "LEARNDO 2026"
clear = np.full((120, 600), 255, np.uint8)
cv2.putText(clear, reference, (24, 80), cv2.FONT_HERSHEY_SIMPLEX, 1.7, 0, 3, cv2.LINE_AA)
blurry = cv2.GaussianBlur(clear, (21, 21), 6)
small = cv2.resize(clear, (60, 12), interpolation=cv2.INTER_AREA)
enlarged = cv2.resize(small, (600, 120), interpolation=cv2.INTER_LINEAR)
variants = {"clear": clear, "blurred": blurry, "downsampled": enlarged}

rows = []
for name, image in variants.items():
    prediction = normalize(pytesseract.image_to_string(image, lang="eng", config="--psm 7"))
    errors = edit_distance(reference, prediction)
    rows.append({"variant": name, "reference": reference, "prediction": prediction,
                 "errors": errors, "cer": cer(reference, prediction)})
    cv2.imwrite(str(output / f"{name}.png"), image)
    print(name, repr(prediction), f"CER={rows[-1]['cer']:.3f}")
with (output / "results.csv").open("w", newline="", encoding="utf-8-sig") as stream:
    writer = csv.DictWriter(stream, fieldnames=list(rows[0]))
    writer.writeheader()
    writer.writerows(rows)
fig, axes = plt.subplots(3, 1, figsize=(10, 5))
for ax, (name, image), row in zip(axes, variants.items(), rows):
    ax.imshow(image, cmap="gray", vmin=0, vmax=255)
    ax.set_title(f"{name}: CER={row['cer']:.3f}")
    ax.axis("off")
fig.tight_layout()
fig.savefig(output / "ocr_error_comparison.png", dpi=150)
plt.close(fig)

.\.venv\Scripts\python.exe ocr_error_measurement.py