Tesseract — Klassik va Ishonchli

4-bob 5 kod bloki 659 so'z

Nega Tesseract?

100+ til, jumladan o'zbek (lotin va kirill) va rus. Bepul, ochiq kodli, CPU da tez, offline. Confidence (ishonch darajasi) beradi — production quality gate uchun oltin xususiyat. Kamchiliklari ham bor, lekin boshlang'ich nuqta har doim Tesseract: agar u yetarli bo'lsa, murakkabroq narsa kerak emas.

O'rnatish

# Ubuntu / Debian / WSL
sudo apt install tesseract-ocr

# Til paketlari: rus, o'zbek lotin, o'zbek kirill
sudo apt install tesseract-ocr-rus tesseract-ocr-uzb tesseract-ocr-uzb-cyrl

# Python wrapper
pip install pytesseract pillow

O'rnatilgan tillarni tekshirish:

tesseract --list-langs
# eng, osd, rus, uzb, uzb_cyrl

Asosiy ishlatish

import pytesseract
from PIL import Image

image = Image.open("scanned_page.png")

# Bitta til
text = pytesseract.image_to_string(image, lang="rus")

# Bir nechta til birga — O'zbekiston hujjatlari uchun tipik holat
text = pytesseract.image_to_string(image, lang="rus+uzb")
print(text)

Muhim nuance: lang="rus+uzb" — Tesseract har so'z uchun ikkala til modelini sinab, yaxshisini oladi. Lekin keraksiz til qo'shmang: har qo'shimcha til biroz sekinlashtiradi va adashish ehtimolini oshiradi. Hujjat faqat ruscha bo'lsa — faqat rus.

PSM — Page Segmentation Mode

Tesseract ga "sahifa qanday tuzilgan" degan maslahat. Noto'g'ri PSM — yomon natijaning eng ko'p uchraydigan sababi (preprocessing dan keyin):

# Kvitansiya: qatorlar tartibsiz, ustun tushunchasi yo'q
text = pytesseract.image_to_string(image, lang="rus", config="--psm 11")

# Bitta qatorli yozuv
text = pytesseract.image_to_string(sign, lang="uzb", config="--psm 7")

Amaliy maslahat: natija g'alati chiqsa (so'zlar aralashib ketgan, ustunlar qo'shilib ketgan) — birinchi navbatda PSM ni o'zgartirib ko'ring: 3 → 6 → 11.

Confidence — ishonch darajasi

image_to_string faqat matn beradi. image_to_data esa har so'z uchun koordinata + confidence (0-100) beradi. Bu production tizimning tayanchi:

data = pytesseract.image_to_data(image, lang="rus+uzb",
                                 output_type=pytesseract.Output.DICT)

words, confs = [], []
for word, conf in zip(data["text"], data["conf"]):
    conf = int(float(conf))            # yangi pytesseract "95.2" kabi satr beradi
    if conf > 0 and word.strip():      # -1 = bo'sh blok, tashlab yuboramiz
        words.append(word)
        confs.append(conf)

avg_conf = sum(confs) / max(1, len(confs))
low_words = [w for w, c in zip(words, confs) if c < 60]

print(f"O'rtacha ishonch: {avg_conf:.0f}%")
print(f"Shubhali so'zlar: {low_words[:10]}")

Confidence bilan nima qilamiz (8-bobda batafsil):

• avg_conf > 80 — natijaga ishonamiz, to'g'ridan-to'g'ri RAG ga
• 60-80 — ishlatamiz, lekin metadata ga belgilab qo'yamiz
• < 60 — VLM bilan qayta o'qiymiz yoki manual review ga

Kuchli va zaif tomonlari

Xulosa