Kalit nuqtalar: burilgan va o'lchami o'zgargan obyektni topish

rasm = cv2.imread(str(RASM / "qurilish_kichik.jpg"))
gray = cv2.cvtColor(rasm, cv2.COLOR_BGR2GRAY)
orb = cv2.ORB_create(nfeatures=1000)
nuqtalar, tavsif = orb.detectAndCompute(gray, None)
print(f"topilgan kalit nuqtalar: {len(nuqtalar)}")
print(f"deskriptor matritsasi: {tavsif.shape} | turi: {tavsif.dtype}")
print(f"birinchi nuqta — joyi: ({nuqtalar[0].pt[0]:.0f}, {nuqtalar[0].pt[1]:.0f}), "
      f"o'lchami: {nuqtalar[0].size:.1f}, burchagi: {nuqtalar[0].angle:.0f}°")

belgi = cv2.drawKeypoints(rasm, nuqtalar[:300], None, color=(0, 255, 0),
                          flags=cv2.DRAW_MATCHES_FLAGS_DRAW_RICH_KEYPOINTS)
ko_rsat(belgi, nom="orb_nuqtalar.png", kenglik=10,
        izoh="Har doira — bitta kalit nuqta. Doira o'lchami masshtabni, chiziq esa burchakni ko'rsatadi.")

sohalar = {"osmon (tekis)": gray[0:120, 300:600], "bino (naqshli)": gray[150:400, 120:420],
           "yer (o'rtacha)": gray[420:540, 100:500]}
show(pd.DataFrame([{"soha": n, "o'lcham": f"{s.shape[1]}x{s.shape[0]}",
                    "kalit nuqtalar": len(orb.detect(s, None)),
                    "1000 pikselga": round(1000 * len(orb.detect(s, None)) / s.size, 2)}
                   for n, s in sohalar.items()]))

M = cv2.getRotationMatrix2D((gray.shape[1] / 2, gray.shape[0] / 2), 30, 0.8)
ozgargan = cv2.warpAffine(gray, M, (gray.shape[1], gray.shape[0]), borderValue=128)
n1, d1 = orb.detectAndCompute(gray, None)
n2, d2 = orb.detectAndCompute(ozgargan, None)
bf = cv2.BFMatcher(cv2.NORM_HAMMING, crossCheck=True)
juftlar = sorted(bf.match(d1, d2), key=lambda m: m.distance)
print(f"nuqtalar: {len(n1)} va {len(n2)} | juftlar: {len(juftlar)}")
print(f"eng yaqin juft masofasi: {juftlar[0].distance:.0f} | eng uzog'i: {juftlar[-1].distance:.0f}")
chizma = cv2.drawMatches(gray, n1, ozgargan, n2, juftlar[:40], None,
                         flags=cv2.DrawMatchesFlags_NOT_DRAW_SINGLE_POINTS)
ko_rsat(chizma, nom="juftlar.png", kenglik=14,
        izoh="30° burilgan va 0.8 barobar kichraytirilgan rasm bilan juftlash. Chiziqlar parallel bo'lsa — to'g'ri juftlar.")

bf2 = cv2.BFMatcher(cv2.NORM_HAMMING)
xom = bf2.knnMatch(d1, d2, k=2)
qatorlar = []
for nisbat in (0.6, 0.7, 0.75, 0.8, 0.9, 1.0):
    yaxshi = [a for a, b in xom if a.distance < nisbat * b.distance]
    qatorlar.append({"ratio": nisbat, "qolgan juftlar": len(yaxshi),
                     "ulush %": round(100 * len(yaxshi) / len(xom), 1)})
show(pd.DataFrame(qatorlar))
yaxshi = [a for a, b in xom if a.distance < 0.75 * b.distance]

src = np.float32([n1[m.queryIdx].pt for m in yaxshi]).reshape(-1, 1, 2)
dst = np.float32([n2[m.trainIdx].pt for m in yaxshi]).reshape(-1, 1, 2)
H, maska = cv2.findHomography(src, dst, cv2.RANSAC, 5.0)
ichkilar = int(maska.sum())
print(f"juftlar: {len(yaxshi)} | RANSAC ichkilari (inliers): {ichkilar} ({100 * ichkilar / len(yaxshi):.0f}%)")
print("topilgan homografiya:\n", np.round(H, 3))

HAQIQAT = np.vstack([M, [0, 0, 1]])                     # biz qo'llagan burilish+masshtab matritsasi
burchaklar = np.float32([[0, 0], [gray.shape[1], 0], [gray.shape[1], gray.shape[0]], [0, gray.shape[0]]]).reshape(-1, 1, 2)
topilgan_b = cv2.perspectiveTransform(burchaklar, H).reshape(4, 2)
haqiqiy_b = cv2.perspectiveTransform(burchaklar, HAQIQAT).reshape(4, 2)
xato = np.linalg.norm(topilgan_b - haqiqiy_b, axis=1)
show(pd.DataFrame([{"burchak": i, "topilgan": f"({t[0]:.0f}, {t[1]:.0f})",
                    "haqiqiy": f"({q[0]:.0f}, {q[1]:.0f})", "xato, piksel": round(float(e), 1)}
                   for i, (t, q, e) in enumerate(zip(topilgan_b, haqiqiy_b, xato))]))

def orb_ichkilar(o1, o2):
    """Ikki rasm orasidagi RANSAC ichkilari soni."""
    a1, t1 = orb.detectAndCompute(o1, None)
    a2, t2 = orb.detectAndCompute(o2, None)
    if t1 is None or t2 is None or len(t1) < 4 or len(t2) < 4:
        return 0
    jufts = bf2.knnMatch(t1, t2, k=2)
    yax = [p[0] for p in jufts if len(p) == 2 and p[0].distance < 0.75 * p[1].distance]
    if len(yax) < 4:
        return 0
    s = np.float32([a1[m.queryIdx].pt for m in yax]).reshape(-1, 1, 2)
    d = np.float32([a2[m.trainIdx].pt for m in yax]).reshape(-1, 1, 2)
    Hm, mk = cv2.findHomography(s, d, cv2.RANSAC, 5.0)
    return int(mk.sum()) if mk is not None else 0

def shablon_ball(katta, shab):
    return round(float(cv2.minMaxLoc(cv2.matchTemplate(katta, shab, cv2.TM_CCOEFF_NORMED))[1]), 3)

shab = gray[150:400, 120:420]                            # binoning bir bo'lagi — shablon
qatorlar = []
for burchak in (0, 5, 10, 20, 45, 90):
    Mr = cv2.getRotationMatrix2D((gray.shape[1] / 2, gray.shape[0] / 2), burchak, 1.0)
    o = cv2.warpAffine(gray, Mr, (gray.shape[1], gray.shape[0]), borderValue=128)
    qatorlar.append({"burilish": f"{burchak}°", "ORB ichkilari": orb_ichkilar(gray, o),
                     "matchTemplate balli": shablon_ball(o, shab)})
show(pd.DataFrame(qatorlar))

qatorlar = []
for koef in (0.5, 0.7, 1.0, 1.4, 2.0):
    o = cv2.resize(gray, None, fx=koef, fy=koef, interpolation=cv2.INTER_AREA)
    sigadi = o.shape[0] >= shab.shape[0] and o.shape[1] >= shab.shape[1]
    qatorlar.append({"o'lcham": f"x{koef}", "ORB ichkilari": orb_ichkilar(gray, o),
                     "matchTemplate balli": shablon_ball(o, shab) if sigadi else None})
show(pd.DataFrame(qatorlar))

import time
usullar = {"ORB": cv2.ORB_create(nfeatures=1000), "SIFT": cv2.SIFT_create(nfeatures=1000),
           "AKAZE": cv2.AKAZE_create(), "BRISK": cv2.BRISK_create()}
Mr = cv2.getRotationMatrix2D((gray.shape[1] / 2, gray.shape[0] / 2), 30, 0.8)
burilgan = cv2.warpAffine(gray, Mr, (gray.shape[1], gray.shape[0]), borderValue=128)
qatorlar = []
for nom, det in usullar.items():
    t0 = time.perf_counter()
    k1, v1 = det.detectAndCompute(gray, None)
    k2, v2 = det.detectAndCompute(burilgan, None)
    ms = (time.perf_counter() - t0) * 1000
    norma = cv2.NORM_L2 if nom == "SIFT" else cv2.NORM_HAMMING
    m = cv2.BFMatcher(norma).knnMatch(v1, v2, k=2)
    yax = [a for a, b in m if a.distance < 0.75 * b.distance]
    ichki = 0
    if len(yax) >= 4:
        s = np.float32([k1[x.queryIdx].pt for x in yax]).reshape(-1, 1, 2)
        d = np.float32([k2[x.trainIdx].pt for x in yax]).reshape(-1, 1, 2)
        _, mk = cv2.findHomography(s, d, cv2.RANSAC, 5.0)
        ichki = int(mk.sum()) if mk is not None else 0
    qatorlar.append({"usul": nom, "nuqtalar": len(k1), "deskriptor": f"{v1.shape[1]} x {v1.dtype}",
                     "yaxshi juftlar": len(yax), "ichkilar": ichki,
                     "ichki ulushi %": round(100 * ichki / max(len(yax), 1), 1), "vaqt, ms": round(ms)})
show(pd.DataFrame(qatorlar))