Yuz aniqlash: Haar kaskadi va YuNet

MOT = DATA.parent.parent / "YOLO_Tracking_Kitob" / "datasets" / "MOT17" / "train" / "MOT17-09-FRCNN"
OGIRLIK = DATA.parent.parent / "YOLO_Tracking_Kitob" / "weights" / "face_detection_yunet_2023mar.onnx"
gt = np.loadtxt(MOT / "gt" / "gt.txt", delimiter=",")
piyoda = gt[(gt[:, 7] == 1) & (gt[:, 6] == 1)]                       # sinf 1 = odam, 6-ustun = e'tiborga olish
tanlov = piyoda[(piyoda[:, 8] >= 0.8) & (piyoda[:, 0] % 20 == 0)]    # ko'rinishi 0.8 dan yuqori, har 20-kadr
tanlov = tanlov[(tanlov[:, 2] >= 0) & (tanlov[:, 3] >= 0) &
                (tanlov[:, 2] + tanlov[:, 4] <= 1920) & (tanlov[:, 3] + tanlov[:, 5] <= 1080)]
print(f"gt.txt qatorlari: {len(gt)} | piyodalar: {len(piyoda)} | sinovga tanlangan: {len(tanlov)}")
print(f"kadrlar: {len(np.unique(tanlov[:, 0]))} ta")
o_w, o_h = tanlov[:, 4], tanlov[:, 5]
print(f"odam qutisi o'lchami — o'rtacha {o_w.mean():.0f}x{o_h.mean():.0f} px, "
      f"eng kichigi {o_w.min():.0f}x{o_h.min():.0f}")

kadr = cv2.imread(str(MOT / "img1" / f"{int(tanlov[0, 0]):06d}.jpg"))
namuna = kadr.copy()
for g in tanlov[tanlov[:, 0] == tanlov[0, 0]]:
    x, y, w, h = g[2:6].astype(int)
    cv2.rectangle(namuna, (x, y), (x + w, y + h), (0, 255, 0), 2)
ko_rsat(cv2.resize(namuna, (960, 540)), nom="mot_kadr.png", kenglik=11,
        izoh="MOT17 kadri: yashil qutilar — qo'lda belgilangan odamlar. Har biridan yuz qidiramiz.")

import time
HAAR = cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
haar = cv2.CascadeClassifier(HAAR)
print("kaskad yuklandimi:", not haar.empty())

def haar_topish(bolaklar, scaleFactor=1.1, minNeighbors=5, minSize=(20, 20)):
    topildi, vaqtlar = 0, []
    for b in bolaklar:
        g = cv2.cvtColor(b, cv2.COLOR_BGR2GRAY)
        t0 = time.perf_counter()
        yuzlar = haar.detectMultiScale(g, scaleFactor=scaleFactor, minNeighbors=minNeighbors, minSize=minSize)
        vaqtlar.append(time.perf_counter() - t0)
        if len(yuzlar) > 0:
            topildi += 1
    return topildi, float(np.mean(vaqtlar)) * 1000

bolaklar = []
for k in np.unique(tanlov[:, 0]).astype(int):
    kadr = cv2.imread(str(MOT / "img1" / f"{k:06d}.jpg"))
    for g in tanlov[tanlov[:, 0] == k]:
        x, y, w, h = g[2:6].astype(int)
        bolaklar.append(kadr[y:y + h, x:x + w])
print(f"kesib olingan odam bo'laklari: {len(bolaklar)}")

haar_topildi, haar_ms_bolak = haar_topish(bolaklar)
print(f"Haar: {haar_topildi} / {len(bolaklar)} bo'lakda yuz topdi "
      f"({100 * haar_topildi / len(bolaklar):.1f}%), bo'lagiga {haar_ms_bolak:.1f} ms")

qatorlar = []
for sf in (1.05, 1.1, 1.3):
    for mn in (3, 5, 8):
        t, t_ms = haar_topish(bolaklar, scaleFactor=sf, minNeighbors=mn)
        qatorlar.append({"scaleFactor": sf, "minNeighbors": mn, "topilgan": t,
                         "topish %": round(100 * t / len(bolaklar), 1), "bo'lagiga, ms": round(t_ms, 1)})
show(pd.DataFrame(qatorlar))

print("model fayl:", OGIRLIK.name, f"{OGIRLIK.stat().st_size / 1e6:.2f} MB")
yunet = cv2.FaceDetectorYN.create(str(OGIRLIK), "", (320, 320), 0.7, 0.3, 5000)

def yunet_topish(bolaklar, ishonch=0.7):
    d = cv2.FaceDetectorYN.create(str(OGIRLIK), "", (320, 320), ishonch, 0.3, 5000)
    topildi, vaqtlar, ballar = 0, [], []
    for b in bolaklar:
        h, w = b.shape[:2]
        d.setInputSize((w, h))
        t0 = time.perf_counter()
        _, F = d.detect(b)
        vaqtlar.append(time.perf_counter() - t0)
        if F is not None and len(F) > 0:
            topildi += 1
            ballar.append(float(F[0][-1]))
    return topildi, float(np.mean(vaqtlar)) * 1000, ballar

t_y, ms_y, ballar = yunet_topish(bolaklar)
print(f"YuNet: {t_y} / {len(bolaklar)} bo'lakda yuz topdi ({100 * t_y / len(bolaklar):.1f}%), "
      f"bo'lagiga {ms_y:.1f} ms")
print(f"ishonch ballari — o'rtacha {np.mean(ballar):.3f}, eng kichigi {min(ballar):.3f}")

qatorlar = [{"detektor": "Haar (default)", "topilgan": haar_topildi,
             "topish %": round(100 * haar_topildi / len(bolaklar), 1),
             "bo'lagiga, ms": round(haar_ms_bolak, 1)}]
for ishonch in (0.5, 0.7, 0.9):
    t_i, ms_i, _ = yunet_topish(bolaklar, ishonch)
    qatorlar.append({"detektor": f"YuNet (ishonch {ishonch})", "topilgan": t_i,
                     "topish %": round(100 * t_i / len(bolaklar), 1), "bo'lagiga, ms": round(ms_i, 1)})
show(pd.DataFrame(qatorlar))

olchamlar = np.array([b.shape[0] for b in bolaklar])
guruhlar = [("kichik (< 200 px)", olchamlar < 200), ("o'rta (200–400)", (olchamlar >= 200) & (olchamlar < 400)),
            ("katta (>= 400)", olchamlar >= 400)]
qatorlar = []
for nom, m in guruhlar:
    guruh = [b for b, bor in zip(bolaklar, m) if bor]
    if not guruh:
        continue
    h_t, _ = haar_topish(guruh)
    y_t, _, _ = yunet_topish(guruh)
    qatorlar.append({"odam balandligi": nom, "bo'laklar": len(guruh),
                     "Haar topish %": round(100 * h_t / len(guruh), 1),
                     "YuNet topish %": round(100 * y_t / len(guruh), 1)})
show(pd.DataFrame(qatorlar))

yunet.setInputSize((320, 320))
namunalar, chizmalar = [], []
for b in bolaklar:
    h, w = b.shape[:2]
    if h < 200:
        continue
    yunet.setInputSize((w, h))
    _, F = yunet.detect(b)
    if F is None or len(F) == 0:
        continue
    f = F[0]
    x, y, bw, bh = f[:4].astype(int)
    ch = b.copy()
    cv2.rectangle(ch, (x, y), (x + bw, y + bh), (0, 255, 0), 2)
    for i, rang in enumerate([(255, 0, 0), (0, 0, 255), (0, 255, 255), (255, 0, 255), (0, 128, 255)]):
        cv2.circle(ch, (int(f[4 + 2 * i]), int(f[5 + 2 * i])), 3, rang, -1)
    chizmalar.append(cv2.resize(ch, (140, 320)))
    if len(chizmalar) >= 4:
        break
print(f"kalit nuqtali namunalar: {len(chizmalar)}")
ko_rsat(*chizmalar, nom="yunet_nuqtalar.png", kenglik=10,
        izoh="YuNet ning 5 nuqtasi: ko'zlar, burun, og'iz burchaklari. Yuzni to'g'rilash uchun shular kerak.")

kadr = cv2.imread(str(MOT / "img1" / f"{int(np.unique(tanlov[:, 0])[2]):06d}.jpg"))
h, w = kadr.shape[:2]
t0 = time.perf_counter()
g = cv2.cvtColor(kadr, cv2.COLOR_BGR2GRAY)
haar_yuzlar = haar.detectMultiScale(g, 1.1, 5, minSize=(20, 20))
haar_ms = (time.perf_counter() - t0) * 1000

d = cv2.FaceDetectorYN.create(str(OGIRLIK), "", (w, h), 0.7, 0.3, 5000)
t0 = time.perf_counter()
_, F = d.detect(kadr)
yunet_ms = (time.perf_counter() - t0) * 1000
n_yunet = 0 if F is None else len(F)
odamlar_shu_kadrda = int((tanlov[:, 0] == np.unique(tanlov[:, 0])[2]).sum())
show(pd.DataFrame([
    {"detektor": "Haar", "topilgan yuzlar": len(haar_yuzlar), "kadrga, ms": round(haar_ms, 1)},
    {"detektor": "YuNet", "topilgan yuzlar": n_yunet, "kadrga, ms": round(yunet_ms, 1)},
]))
print(f"kadr o'lchami: {w}x{h} | bu kadrda belgilangan odamlar: {odamlar_shu_kadrda}")
print("eslatma: butun kadrda Haar har piksel pozitsiyasini tekshiradi — shuning uchun sekin")

belgi = kadr.copy()
for (x, y, bw, bh) in haar_yuzlar:
    cv2.rectangle(belgi, (x, y), (x + bw, y + bh), (0, 0, 255), 3)
if F is not None:
    for f in F:
        x, y, bw, bh = f[:4].astype(int)
        cv2.rectangle(belgi, (x, y), (x + bw, y + bh), (0, 255, 0), 3)
ko_rsat(cv2.resize(belgi, (960, 540)), nom="butun_kadr_yuz.png", kenglik=11,
        izoh="Qizil — Haar, yashil — YuNet. Butun kadrda ishlatilganda farq yaqqol ko'rinadi.")