dyagnosys-free / tabs /deception_detection.py
vitorcalvi's picture
1
18c46ab
# tabs/deception_detection.py
import gradio as gr
import cv2
import numpy as np
import matplotlib.pyplot as plt
from scipy.signal import butter, filtfilt, find_peaks
from typing import Tuple, Optional, Dict
import logging
from dataclasses import dataclass
from enum import Enum
import librosa
import moviepy.editor as mp
import os
import tempfile
import torch
import torch.nn as nn
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import mediapipe as mp_mediapipe
import re
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Define Enums and DataClasses
class DeceptionLevel(Enum):
LOW = 'Low'
MODERATE = 'Moderate'
HIGH = 'High'
@dataclass
class Metric:
name: str
threshold: float
value: float = 0.0
detected: bool = False
def analyze(self, new_value: float):
self.value = new_value
self.detected = self.value > self.threshold
class SignalProcessor:
def __init__(self, fs: float):
self.fs = fs # Sampling frequency
def bandpass_filter(self, data: np.ndarray, lowcut: float = 0.75, highcut: float = 3.0) -> np.ndarray:
"""Apply bandpass filter to signal."""
nyq = 0.5 * self.fs
low = lowcut / nyq
high = highcut / nyq
b, a = butter(2, [low, high], btype='band')
filtered = filtfilt(b, a, data)
logger.debug("Applied bandpass filter.")
return filtered
def find_peaks_in_signal(self, signal: np.ndarray) -> np.ndarray:
"""Find peaks in the signal."""
min_distance = int(60 / 180 * self.fs) # At least 60 BPM (180 BPM max)
peaks, _ = find_peaks(signal, distance=min_distance)
logger.debug(f"Detected {len(peaks)} peaks in the signal.")
return peaks
class DeceptionAnalyzer:
def __init__(self):
self.metrics = {
"HRV Suppression": Metric("HRV Suppression", threshold=30.0),
"Heart Rate Elevation": Metric("Heart Rate Elevation", threshold=100.0),
"Rhythm Irregularity": Metric("Rhythm Irregularity", threshold=0.1),
"Blink Rate": Metric("Blink Rate", threshold=25.0),
"Head Movements": Metric("Head Movements", threshold=10.0),
"Speech Stress": Metric("Speech Stress", threshold=0.5),
"Speech Pitch Variation": Metric("Speech Pitch Variation", threshold=50.0),
"Pauses and Hesitations": Metric("Pauses and Hesitations", threshold=2.0),
"Filler Words": Metric("Filler Words", threshold=5.0),
}
def analyze_signals(self, heart_rate: np.ndarray, rr_intervals: np.ndarray, hrv_rmssd: float,
speech_features: Dict[str, float], facial_features: Dict[str, float]) -> Tuple[Dict[str, Dict], float, DeceptionLevel]:
"""
Analyze the extracted signals and compute deception probability.
"""
# Analyze HRV Suppression
self.metrics["HRV Suppression"].analyze(hrv_rmssd)
# Analyze Heart Rate Elevation
avg_heart_rate = np.mean(heart_rate)
self.metrics["Heart Rate Elevation"].analyze(avg_heart_rate)
# Analyze Rhythm Irregularity
rhythm_irregularity = np.std(rr_intervals) / np.mean(rr_intervals)
self.metrics["Rhythm Irregularity"].analyze(rhythm_irregularity)
# Analyze Speech Features
for key in ["Speech Stress", "Speech Pitch Variation", "Pauses and Hesitations", "Filler Words"]:
if key in speech_features:
self.metrics[key].analyze(speech_features[key])
# Analyze Facial Features
# Placeholder values; in actual implementation, replace with real values
self.metrics["Blink Rate"].analyze(facial_features.get("Blink Rate", 0))
self.metrics["Head Movements"].analyze(facial_features.get("Head Movements", 0))
# Calculate deception probability
detected_indicators = sum(1 for m in self.metrics.values() if m.detected)
total_indicators = len(self.metrics)
probability = (detected_indicators / total_indicators) * 100
# Determine deception level
if probability < 30:
level = DeceptionLevel.LOW
elif probability < 70:
level = DeceptionLevel.MODERATE
else:
level = DeceptionLevel.HIGH
# Prepare metrics for visualization
metrics_data = {name: {
"value": m.value,
"threshold": m.threshold,
"detected": m.detected
} for name, m in self.metrics.items()}
return metrics_data, probability, level
def load_transcription_model(model_name: str) -> Optional[torch.nn.Module]:
"""
Load the speech-to-text transcription model.
"""
try:
model = Wav2Vec2ForCTC.from_pretrained(
model_name,
ignore_mismatched_sizes=True
)
model.eval()
logger.info("Transcription model loaded successfully.")
return model
except Exception as e:
logger.error(f"Error loading transcription model: {e}")
return None
def load_models() -> Dict[str, torch.nn.Module]:
"""
Load all necessary models for the deception detection system.
"""
models_dict = {}
try:
# Load Transcription Model
transcription_model_name = 'facebook/wav2vec2-base-960h'
transcription_model = load_transcription_model(transcription_model_name)
if transcription_model:
models_dict['transcription_model'] = transcription_model
except Exception as e:
logger.error(f"Error loading models: {e}")
return models_dict
def transcribe_audio(audio_path: str, transcription_model: nn.Module) -> str:
"""
Transcribe audio to text using Wav2Vec2 model.
"""
try:
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
y, sr = librosa.load(audio_path, sr=16000)
input_values = tokenizer(y, return_tensors="pt", padding="longest").input_values
with torch.no_grad():
logits = transcription_model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.decode(predicted_ids[0])
# Clean transcription
transcription = transcription.lower()
transcription = re.sub(r'[^a-z\s]', '', transcription)
return transcription
except Exception as e:
logger.error(f"Error transcribing audio: {str(e)}")
return ""
def detect_silence(y: np.ndarray, sr: int, top_db: int = 30) -> float:
"""
Detect total duration of silence in the audio.
"""
try:
intervals = librosa.effects.split(y, top_db=top_db)
silence_duration = 0.0
prev_end = 0
for start, end in intervals:
silence = (start - prev_end) / sr
silence_duration += silence
prev_end = end
# Add silence after the last interval
silence_duration += (len(y) - prev_end) / sr
return silence_duration
except Exception as e:
logger.error(f"Error detecting silence: {str(e)}")
return 0.0
def count_filler_words(transcription: str) -> int:
"""
Count the number of filler words in the transcription.
"""
filler_words_list = ['um', 'uh', 'er', 'ah', 'like', 'you know', 'so']
return sum(transcription.split().count(word) for word in filler_words_list)
def analyze_speech(audio_path: str, transcription_model: nn.Module) -> Dict[str, float]:
"""
Analyze speech from the audio file and extract features.
"""
if not audio_path:
logger.warning("No audio path provided.")
return {}
try:
# Load audio file
y, sr = librosa.load(audio_path, sr=16000) # Ensure consistent sampling rate
logger.info(f"Loaded audio file with sampling rate: {sr} Hz")
# Extract prosodic features
pitches, magnitudes = librosa.piptrack(y=y, sr=sr)
pitch_values = pitches[magnitudes > np.median(magnitudes)]
avg_pitch = np.mean(pitch_values) if len(pitch_values) > 0 else 0.0
pitch_variation = np.std(pitch_values) if len(pitch_values) > 0 else 0.0
# Calculate speech stress based on pitch variation
speech_stress = pitch_variation / (avg_pitch if avg_pitch != 0 else 1)
# Calculate speech rate (words per minute)
transcription = transcribe_audio(audio_path, transcription_model)
words = transcription.split()
duration_minutes = librosa.get_duration(y=y, sr=sr) / 60
speech_rate = len(words) / duration_minutes if duration_minutes > 0 else 0.0
# Detect pauses and hesitations
silence_duration = detect_silence(y, sr)
filler_words = count_filler_words(transcription)
logger.info(f"Speech Analysis - Avg Pitch: {avg_pitch:.2f} Hz, Pitch Variation: {pitch_variation:.2f} Hz")
logger.info(f"Speech Stress Level: {speech_stress:.2f}")
logger.info(f"Speech Rate: {speech_rate:.2f} WPM")
logger.info(f"Silence Duration: {silence_duration:.2f} seconds")
logger.info(f"Filler Words Count: {filler_words}")
# Return extracted features
return {
"Speech Stress": speech_stress,
"Speech Pitch Variation": pitch_variation,
"Pauses and Hesitations": silence_duration,
"Filler Words": filler_words
}
except Exception as e:
logger.error(f"Error analyzing speech: {str(e)}")
return {}
def extract_audio_from_video(video_path: str) -> Optional[str]:
"""
Extract audio from the video file and save it as a temporary WAV file.
"""
if not video_path:
logger.warning("No video path provided for audio extraction.")
return None
try:
video_clip = mp.VideoFileClip(video_path)
if video_clip.audio is None:
logger.warning("No audio track found in the video.")
video_clip.close()
return None
temp_audio_fd, temp_audio_path = tempfile.mkstemp(suffix=".wav")
os.close(temp_audio_fd) # Close the file descriptor
video_clip.audio.write_audiofile(temp_audio_path, logger=None)
video_clip.close()
logger.info(f"Extracted audio to temporary file: {temp_audio_path}")
return temp_audio_path
except Exception as e:
logger.error(f"Error extracting audio from video: {str(e)}")
return None
def detect_blink(face_landmarks, frame: np.ndarray) -> float:
"""
Detect blink rate from facial landmarks.
Placeholder implementation.
"""
# Implement Eye Aspect Ratio (EAR) or other blink detection methods
return np.random.uniform(10, 20) # Example blink rate
def estimate_head_movement(face_landmarks) -> float:
"""
Estimate head movements based on facial landmarks.
Placeholder implementation.
"""
# Implement head pose estimation to detect nods/shakes
return np.random.uniform(5, 15) # Example head movements
def create_visualization(metrics: Dict, probability: float, heart_rate: np.ndarray,
duration: float, level: DeceptionLevel, speech_features: Dict[str, float]) -> plt.Figure:
"""
Create visualization of analysis results.
"""
# Set figure style parameters
plt.style.use('default')
plt.rcParams.update({
'figure.facecolor': 'white',
'axes.facecolor': 'white',
'grid.color': '#E0E0E0',
'grid.linestyle': '-',
'grid.alpha': 0.3,
'font.size': 10,
'axes.labelsize': 10,
'axes.titlesize': 12,
'figure.titlesize': 14,
'font.family': ['DejaVu Sans', 'Arial', 'sans-serif']
})
# Create figure and axes
fig = plt.figure(figsize=(12, 20))
# Create polar plot for deception probability gauge
ax1 = fig.add_subplot(4, 1, 1, projection='polar')
# Create other subplots
ax2 = fig.add_subplot(4, 1, 2)
ax3 = fig.add_subplot(4, 1, 3)
ax4 = fig.add_subplot(4, 1, 4)
# Plot 1: Deception Probability Gauge
# Create gauge plot
theta = np.linspace(0, np.pi, 100)
radius = np.ones(100)
ax1.plot(theta, radius, color='#E0E0E0', linewidth=30, alpha=0.3)
current_angle = (probability / 100) * np.pi
ax1.plot([0, current_angle], [0, 0.7], color='red', linewidth=5)
ax1.set_xticks([])
ax1.set_yticks([])
ax1.set_title(f'Deception Probability: {probability:.1f}% ({level.value})', pad=20, color='#333333')
ax1.set_theta_zero_location('N')
ax1.set_facecolor('white')
ax1.grid(False)
ax1.spines['polar'].set_visible(False)
# Plot 2: Metrics Bar Chart
names = list(metrics.keys())
values = [m["value"] for m in metrics.values()]
thresholds = [m["threshold"] for m in metrics.values()]
detected = [m["detected"] for m in metrics.values()]
x = np.arange(len(names))
width = 0.35
bar_colors = ['#FF6B6B' if d else '#4BB543' for d in detected]
ax2.bar(x - width/2, values, width, label='Current', color=bar_colors)
ax2.bar(x + width/2, thresholds, width, label='Threshold', color='#E0E0E0', alpha=0.7)
ax2.set_ylabel('Value')
ax2.set_title('Physiological, Facial, and Speech Indicators', pad=20)
ax2.set_xticks(x)
ax2.set_xticklabels(names, rotation=45, ha='right')
ax2.grid(True, axis='y', alpha=0.3)
ax2.legend(loc='upper right', framealpha=0.9)
# Plot 3: Heart Rate Over Time
time_axis = np.linspace(0, duration, len(heart_rate))
ax3.plot(time_axis, heart_rate, color='#3498db')
ax3.set_xlabel('Time (s)')
ax3.set_ylabel('Heart Rate (BPM)')
ax3.set_title('Heart Rate Over Time', pad=20)
ax3.grid(True, alpha=0.3)
# Plot 4: Speech Features
pauses = speech_features.get("Pauses and Hesitations", 0)
filler_words = speech_features.get("Filler Words", 0)
labels = ['Pauses (s)', 'Filler Words (count)']
values = [pauses, filler_words]
colors = ['#FFC300', '#FF5733']
ax4.bar(labels, values, color=colors)
ax4.set_ylabel('Count / Duration')
ax4.set_title('Pauses and Hesitations in Speech', pad=20)
ax4.grid(True, axis='y', alpha=0.3)
plt.tight_layout()
return fig
def process_video_and_audio(video_path: str, models: Dict[str, torch.nn.Module]) -> Tuple[Optional[np.ndarray], Optional[plt.Figure]]:
"""
Process video and audio, perform deception analysis.
"""
logger.info("Starting video and audio processing.")
if not video_path:
logger.warning("No video path provided.")
return None, None
try:
# Extract audio from video
audio_path = extract_audio_from_video(video_path)
if not audio_path:
logger.warning("No audio available for speech analysis.")
# Initialize video capture
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
logger.error("Failed to open video file.")
return None, None
fps = cap.get(cv2.CAP_PROP_FPS)
if fps <= 0 or fps != fps:
logger.error("Invalid frame rate detected.")
cap.release()
return None, None
logger.info(f"Video FPS: {fps}")
# Initialize processors
signal_processor = SignalProcessor(fps)
analyzer = DeceptionAnalyzer()
ppg_signal = []
last_frame = None
# Initialize Mediapipe for real-time facial feature extraction
mp_face_mesh = mp_mediapipe.solutions.face_mesh
face_mesh = mp_face_mesh.FaceMesh(static_image_mode=False, max_num_faces=1)
frame_counter = 0
# Process video frames
while True:
ret, frame = cap.read()
if not ret:
break
frame_counter += 1
# Extract PPG signal from green channel
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
green_channel = frame_rgb[:, :, 1]
ppg_signal.append(np.mean(green_channel))
# Extract facial features
results = face_mesh.process(frame_rgb)
if results.multi_face_landmarks:
face_landmarks = results.multi_face_landmarks[0]
# Blink Detection
blink = detect_blink(face_landmarks, frame)
analyzer.metrics["Blink Rate"].analyze(blink)
# Head Movement Detection
head_movement = estimate_head_movement(face_landmarks)
analyzer.metrics["Head Movements"].analyze(head_movement)
else:
analyzer.metrics["Blink Rate"].analyze(0.0)
analyzer.metrics["Head Movements"].analyze(0.0)
# Store last frame
last_frame = cv2.resize(frame_rgb, (320, 240))
# Optional: Log progress every 100 frames
if frame_counter % 100 == 0:
logger.info(f"Processed {frame_counter} frames.")
cap.release()
face_mesh.close()
logger.info(f"Total frames processed: {frame_counter}")
if not ppg_signal or last_frame is None:
logger.error("No PPG signal extracted or last frame missing.")
return last_frame, None
# Convert PPG signal to numpy array
ppg_signal = np.array(ppg_signal)
logger.debug("PPG signal extracted.")
# Apply bandpass filter
filtered_signal = signal_processor.bandpass_filter(ppg_signal)
logger.debug("Filtered PPG signal.")
# Find peaks in the filtered signal
peaks = signal_processor.find_peaks_in_signal(filtered_signal)
if len(peaks) < 2:
logger.warning("Insufficient peaks detected. Signal quality may be poor.")
return last_frame, None # Return last_frame but no analysis
# Calculate RR intervals in milliseconds
rr_intervals = np.diff(peaks) / fps * 1000 # ms
heart_rate = 60 * fps / np.diff(peaks) # BPM
if len(rr_intervals) == 0 or len(heart_rate) == 0:
logger.error("Failed to calculate RR intervals or heart rate.")
return last_frame, None
# Calculate RMSSD (Root Mean Square of Successive Differences)
hrv_rmssd = np.sqrt(np.mean(np.diff(rr_intervals) ** 2))
logger.debug(f"Calculated RMSSD: {hrv_rmssd:.2f} ms")
# Analyze speech
if audio_path and 'transcription_model' in models:
speech_features = analyze_speech(audio_path, models['transcription_model'])
else:
speech_features = {}
# Analyze signals
metrics, probability, level = analyzer.analyze_signals(
heart_rate, rr_intervals, hrv_rmssd, speech_features,
{}
)
# Create visualization
duration = len(ppg_signal) / fps # seconds
fig = create_visualization(
metrics, probability, heart_rate,
duration, level, speech_features
)
# Clean up temporary audio file if it was extracted
if audio_path and os.path.exists(audio_path):
try:
os.remove(audio_path)
logger.info(f"Deleted temporary audio file: {audio_path}")
except Exception as e:
logger.error(f"Error deleting temporary audio file: {str(e)}")
logger.info("Video and audio processing completed successfully.")
return last_frame, fig
except Exception as e:
logger.error(f"Error processing video and audio: {str(e)}")
return None, None
def create_deception_detection_tab(models: Dict[str, torch.nn.Module]) -> gr.Blocks:
"""
Create the deception detection interface tab using Gradio.
"""
def analyze(video):
try:
if video is None:
return None, None
video_path = video
logger.info(f"Received video for analysis: {video_path}")
if not os.path.exists(video_path):
logger.error("Video file does not exist.")
return None, None
last_frame, fig = process_video_and_audio(video_path, models)
if fig:
return last_frame, fig
else:
return last_frame, None
except Exception as e:
logger.error(f"Error in analyze function: {str(e)}")
return None, None
with gr.Blocks() as deception_interface:
with gr.Row():
with gr.Column(scale=1):
input_video = gr.Video()
gr.Examples(["./assets/videos/fitness.mp4", "./assets/videos/vladirmir.mp4", "./assets/videos/lula.mp4"], inputs=[input_video])
gr.Markdown("""
### Deception Level Analysis
This analysis evaluates physiological, facial, and speech indicators
that may suggest deceptive behavior.
**Physiological Indicators:**
- β—‡ HRV Suppression
- β—‡ Heart Rate Elevation
- β—‡ Rhythm Irregularity
**Facial Indicators:**
- β—‡ Blink Rate
- β—‡ Head Movements
**Speech Indicators:**
- β—‡ Speech Stress
- β—‡ Speech Pitch Variation
- β—‡ Pauses and Hesitations
- β—‡ Filler Words
**Interpretation:**
- **Low (0-30%):** Minimal indicators
- **Moderate (30-70%):** Some indicators
- **High (>70%):** Strong indicators
**Important Note:**
This analysis is for research purposes only.
Results should not be used as definitive proof
of deception or truthfulness.
""")
with gr.Column(scale=2):
output_frame = gr.Image(label="Last Frame of Video", height=240)
analysis_plot = gr.Plot(label="Deception Analysis")
# Configure automatic analysis upon video upload
input_video.change(
fn=analyze,
inputs=[input_video],
outputs=[output_frame, analysis_plot]
)
return deception_interface