Spaces:
Runtime error
Runtime error
File size: 8,079 Bytes
fba30db | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 | """Phase 17.1 β Temporal Consistency Module.
Analyses optical flow variance, luminance flicker, and blink timing across
sampled video frames to produce a temporal_score (0β100, higher = more natural).
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import List, Tuple
import cv2
import numpy as np
from loguru import logger
@dataclass
class TemporalAnalysis:
temporal_score: float # 0β100, higher = more natural / authentic
optical_flow_variance: float # mean inter-frame flow-magnitude variance
flicker_score: float # 0β100 (high = suspicious micro-flicker)
blink_rate_anomaly: bool # True when blink timing is unnatural
blink_intervals: List[float] = field(default_factory=list)
diagnostics: dict = field(default_factory=dict)
# ---------------------------------------------------------------------------
# Optical-flow variance
# ---------------------------------------------------------------------------
def _compute_optical_flow_variance(frames_bgr: List[np.ndarray]) -> float:
"""Mean variance of inter-frame optical-flow magnitudes.
Real videos show consistent, smooth motion; deepfake temporal inconsistencies
appear as irregular per-frame flow jumps.
"""
if len(frames_bgr) < 2:
return 0.0
flow_mags: List[float] = []
for i in range(len(frames_bgr) - 1):
prev_gray = cv2.cvtColor(frames_bgr[i], cv2.COLOR_BGR2GRAY)
curr_gray = cv2.cvtColor(frames_bgr[i + 1], cv2.COLOR_BGR2GRAY)
h, w = prev_gray.shape
scale = min(1.0, 320.0 / max(h, w, 1))
if scale < 1.0:
dsize = (max(1, int(w * scale)), max(1, int(h * scale)))
prev_gray = cv2.resize(prev_gray, dsize)
curr_gray = cv2.resize(curr_gray, dsize)
flow = cv2.calcOpticalFlowFarneback(
prev_gray, curr_gray, None,
pyr_scale=0.5, levels=3, winsize=15,
iterations=3, poly_n=5, poly_sigma=1.2, flags=0,
)
mag, _ = cv2.cartToPolar(flow[..., 0], flow[..., 1])
flow_mags.append(float(np.mean(mag)))
return float(np.var(flow_mags)) if flow_mags else 0.0
# ---------------------------------------------------------------------------
# Luminance flicker
# ---------------------------------------------------------------------------
def _compute_flicker_score(frames_bgr: List[np.ndarray]) -> float:
"""Flicker score 0β100 derived from inter-frame luminance variance.
Deepfake GAN generators introduce subtle luminance micro-flicker that
manifests as high variance in the difference sequence.
"""
if len(frames_bgr) < 2:
return 0.0
mean_lums = [
float(np.mean(cv2.cvtColor(f, cv2.COLOR_BGR2GRAY)))
for f in frames_bgr
]
diffs = [abs(mean_lums[i + 1] - mean_lums[i]) for i in range(len(mean_lums) - 1)]
if not diffs:
return 0.0
mean_diff = float(np.mean(diffs))
std_diff = float(np.std(diffs))
flicker_ratio = std_diff / (mean_diff + 1e-6)
return float(min(100.0, flicker_ratio * 50.0))
# ---------------------------------------------------------------------------
# Blink-rate anomaly (FaceMesh EAR)
# ---------------------------------------------------------------------------
def _compute_blink_anomaly(
frames_bgr: List[np.ndarray],
timestamps: List[float],
) -> Tuple[bool, List[float]]:
"""Detect unnatural blink timing using MediaPipe FaceMesh eye-aspect-ratio.
Returns (anomaly_detected, blink_interval_list_seconds).
Natural blink rate: ~15β20/min β intervals ~3β4 s.
Anomalies: perfectly regular cadence (std < 0.05 s) or rate > 2/s.
"""
try:
import mediapipe as mp
mp_face_mesh = mp.solutions.face_mesh
except ImportError:
return False, []
# Landmark indices for left eye (vertical & horizontal pairs)
EYE_V = (159, 145)
EYE_H = (33, 133)
BLINK_THRESH = 0.25
ear_seq: List[Tuple[float, float]] = []
with mp_face_mesh.FaceMesh(
static_image_mode=True,
max_num_faces=1,
refine_landmarks=True,
min_detection_confidence=0.5,
) as mesh:
for frame_bgr, ts in zip(frames_bgr, timestamps):
rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
res = mesh.process(rgb)
if not (res and res.multi_face_landmarks):
continue
lm = res.multi_face_landmarks[0].landmark
h, w = frame_bgr.shape[:2]
def pt(idx: int) -> np.ndarray:
return np.array([lm[idx].x * w, lm[idx].y * h])
v = float(np.linalg.norm(pt(EYE_V[0]) - pt(EYE_V[1])))
h_dist = float(np.linalg.norm(pt(EYE_H[0]) - pt(EYE_H[1])))
ear = v / (h_dist + 1e-6)
ear_seq.append((ts, ear))
if len(ear_seq) < 3:
return False, []
blink_times: List[float] = []
in_blink = False
for ts, ear in ear_seq:
if ear < BLINK_THRESH and not in_blink:
blink_times.append(ts)
in_blink = True
elif ear >= BLINK_THRESH:
in_blink = False
if len(blink_times) < 2:
return False, []
intervals = [
round(blink_times[i + 1] - blink_times[i], 3)
for i in range(len(blink_times) - 1)
]
mean_iv = float(np.mean(intervals))
std_iv = float(np.std(intervals))
anomaly = (std_iv < 0.05 and len(intervals) > 2) or mean_iv < 0.5
return bool(anomaly), intervals
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def compute_temporal_score(
frames_bgr: List[np.ndarray],
timestamps: List[float],
) -> TemporalAnalysis:
"""Compute temporal consistency for a list of BGR video frames.
Args:
frames_bgr: BGR numpy arrays in temporal order.
timestamps: Corresponding timestamps in seconds.
Returns:
TemporalAnalysis with temporal_score 0β100 (higher = more authentic).
"""
if len(frames_bgr) < 2:
return TemporalAnalysis(
temporal_score=50.0,
optical_flow_variance=0.0,
flicker_score=0.0,
blink_rate_anomaly=False,
diagnostics={"frames_analyzed": len(frames_bgr)},
)
flow_var = 0.0
try:
flow_var = _compute_optical_flow_variance(frames_bgr)
except Exception as exc: # noqa: BLE001
logger.warning(f"Optical flow failed: {exc}")
flicker = 0.0
try:
flicker = _compute_flicker_score(frames_bgr)
except Exception as exc: # noqa: BLE001
logger.warning(f"Flicker score failed: {exc}")
blink_anomaly, blink_intervals = False, []
try:
blink_anomaly, blink_intervals = _compute_blink_anomaly(frames_bgr, timestamps)
except Exception as exc: # noqa: BLE001
logger.warning(f"Blink rate analysis failed: {exc}")
# Score composition
# flow_var: real ~0β2; deepfake inconsistencies push higher β penalise
flow_auth = max(0.0, 100.0 - min(100.0, flow_var * 15.0))
flicker_auth = 100.0 - flicker
blink_penalty = 20.0 if blink_anomaly else 0.0
# Weights: 50% flow, 40% flicker, 10% blink
raw = 0.50 * flow_auth + 0.40 * flicker_auth + 0.10 * (100.0 - blink_penalty)
temporal_score = float(max(0.0, min(100.0, raw)))
logger.info(
f"Temporal: flow_var={flow_var:.4f} flicker={flicker:.1f} "
f"blink_anomaly={blink_anomaly} β temporal_score={temporal_score:.1f}"
)
return TemporalAnalysis(
temporal_score=round(temporal_score, 2),
optical_flow_variance=round(flow_var, 4),
flicker_score=round(flicker, 2),
blink_rate_anomaly=blink_anomaly,
blink_intervals=blink_intervals,
diagnostics={
"flow_component": round(flow_auth, 1),
"flicker_component": round(flicker_auth, 1),
"frames_analyzed": len(frames_bgr),
},
)
|