Threshold: rasmni ikki rangga bo'lish

hujjat = cv2.imread(str(RASM / "hujjat.jpg"))
gray = cv2.cvtColor(hujjat, cv2.COLOR_BGR2GRAY)
chap, ong = gray[:, :400], gray[:, 400:]
print(f"chap yarmining o'rtacha yorqinligi: {chap.mean():.0f} | o'ng yarmi: {ong.mean():.0f}")
print(f"butun rasm: min {gray.min()}, max {gray.max()}, o'rtacha {gray.mean():.0f}")

natijalar = {}
for chegara in (100, 140, 180):
    _, b = cv2.threshold(gray, chegara, 255, cv2.THRESH_BINARY)
    natijalar[f"chegara {chegara}"] = b
ko_rsat(*natijalar.values(), nom="global_threshold.png", sarlavhalar=list(natijalar),
        izoh="Bitta chegara: past bo'lsa soyadagi joy qorayib ketadi, yuqori bo'lsa yorug' joydagi matn yo'qoladi.")

BURCHAK = np.float32([[150, 90], [720, 170], [650, 940], [90, 830]])     # varaqning 4 burchagi (4-bob)
NISHON = np.float32([[0, 0], [640, 0], [640, 900], [0, 900]])
P = cv2.getPerspectiveTransform(BURCHAK, NISHON)
togri = cv2.warpPerspective(gray, P, (640, 900))                       # varaqni to'g'rilab olamiz

toza = cv2.imread(str(RASM / "hujjat_toza.png"), cv2.IMREAD_GRAYSCALE)  # asl, buzilmagan varaq
_, HAQIQAT = cv2.threshold(toza, 128, 255, cv2.THRESH_BINARY_INV)       # haqiqiy matn piksellari

def sifat(binar):
    """To'g'rilangan varaqdagi natijani haqiqat bilan solishtiradi: matn piksellari bo'yicha F1."""
    topilgan = binar == 0
    haqiqiy = HAQIQAT > 0
    tp = int((topilgan & haqiqiy).sum())
    fp = int((topilgan & ~haqiqiy).sum())
    fn = int((~topilgan & haqiqiy).sum())
    p_ = tp / max(tp + fp, 1)
    r_ = tp / max(tp + fn, 1)
    return round(100 * p_, 1), round(100 * r_, 1), round(100 * 2 * p_ * r_ / max(p_ + r_, 1e-9), 1)

qatorlar = []
for chegara in (100, 140, 180):
    _, b = cv2.threshold(togri, chegara, 255, cv2.THRESH_BINARY)
    aniq, toliq, f1 = sifat(b)
    qatorlar.append({"usul": f"global {chegara}", "aniqlik %": aniq, "to'liqlik %": toliq, "F1 %": f1})
show(pd.DataFrame(qatorlar))

chegara, otsu = cv2.threshold(togri, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
print("Otsu tanlagan chegara:", int(chegara), "| F1:", sifat(otsu)[2], "%")
_, qolda = cv2.threshold(togri, 140, 255, cv2.THRESH_BINARY)
ko_rsat(otsu, qolda, nom="otsu.png", sarlavhalar=[f"Otsu (chegara {int(chegara)})", "qo'lda 140"],
        izoh="Otsu chegarani o'zi topadi, lekin u ham bitta chegara — notekis yorug'lik muammosini yechmaydi.")

mean_t = cv2.adaptiveThreshold(togri, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 31, 10)
gauss_t = cv2.adaptiveThreshold(togri, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 31, 10)
ko_rsat(otsu, mean_t, gauss_t, nom="adaptive.png",
        sarlavhalar=["Otsu (global)", "adaptive MEAN", "adaptive GAUSSIAN"],
        izoh="Adaptive usulda soyadagi matn ham ajraladi: to'liqlik 99.6% (Otsu da 90.9%).")
show(pd.DataFrame([{"usul": n, "aniqlik %": sifat(b)[0], "to'liqlik %": sifat(b)[1], "F1 %": sifat(b)[2]}
                   for n, b in [("global 140", qolda), ("Otsu", otsu),
                                ("adaptive MEAN", mean_t), ("adaptive GAUSSIAN", gauss_t)]]))

qatorlar = []
for blok in (11, 31, 61):
    for C in (5, 15):
        b = cv2.adaptiveThreshold(togri, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, blok, C)
        aniq, toliq, f1 = sifat(b)
        qatorlar.append({"blockSize": blok, "C": C, "aniqlik %": aniq, "to'liqlik %": toliq, "F1 %": f1})
show(pd.DataFrame(qatorlar))

turlar = {"BINARY": cv2.THRESH_BINARY, "BINARY_INV": cv2.THRESH_BINARY_INV, "TRUNC": cv2.THRESH_TRUNC,
          "TOZERO": cv2.THRESH_TOZERO}
shakllar = cv2.imread(str(RASM / "shakllar.png"), cv2.IMREAD_GRAYSCALE)
rasmlar = [cv2.threshold(shakllar, 150, 255, t)[1] for t in turlar.values()]
ko_rsat(*rasmlar, nom="thresh_turlari.png", sarlavhalar=list(turlar), kenglik=14,
        izoh="BINARY_INV — teskari (matn oq bo'lib chiqadi, kontur topishda shu kerak bo'ladi).")