liveness_detector.py

"""
Liveness Detection Module
Detects if the face is from a live person or a photo/video spoof
"""

import cv2
import numpy as np
import logging
from typing import Tuple, Optional

logger = logging.getLogger(__name__)


class LivenessDetector:
    """
    Liveness detection using multiple techniques:
    1. Texture analysis (LBP)
    2. Color space analysis
    3. Frequency domain analysis
    4. Eye blink detection (optional)
    """
    
    def __init__(self):
        """Initialize liveness detector"""
        self.lbp_threshold = 0.5
        self.color_threshold = 0.4
        self.frequency_threshold = 0.3
    
    def detect_liveness(self, image: np.ndarray, face: Tuple[int, int, int, int]) -> bool:
        """
        Detect if face is from a live person
        
        Args:
            image: Input image (BGR format)
            face: Face bounding box (x, y, w, h)
        
        Returns:
            True if live, False if spoof detected
        """
        try:
            x, y, w, h = face
            face_roi = image[y:y+h, x:x+w]
            
            if face_roi.size == 0:
                return False
            
            # Run multiple liveness checks
            scores = []
            
            # 1. Texture analysis
            texture_score = self._analyze_texture(face_roi)
            scores.append(texture_score)
            
            # 2. Color space analysis
            color_score = self._analyze_color_space(face_roi)
            scores.append(color_score)
            
            # 3. Frequency domain analysis
            frequency_score = self._analyze_frequency(face_roi)
            scores.append(frequency_score)
            
            # 4. Moiré pattern detection
            moire_score = self._detect_moire_pattern(face_roi)
            scores.append(moire_score)
            
            # Combine scores (weighted average)
            weights = [0.3, 0.25, 0.25, 0.2]
            final_score = sum(s * w for s, w in zip(scores, weights))
            
            # Threshold for liveness
            is_live = final_score > 0.5
            
            logger.debug(f"Liveness scores: texture={texture_score:.3f}, color={color_score:.3f}, "
                        f"frequency={frequency_score:.3f}, moire={moire_score:.3f}, "
                        f"final={final_score:.3f}, live={is_live}")
            
            return is_live
            
        except Exception as e:
            logger.error(f"Liveness detection error: {e}")
            # Default to True to avoid false rejections
            return True
    
    def _analyze_texture(self, face_roi: np.ndarray) -> float:
        """
        Analyze texture using Local Binary Patterns
        Real faces have more complex texture than printed photos
        """
        try:
            gray = cv2.cvtColor(face_roi, cv2.COLOR_BGR2GRAY)
            
            # Compute LBP
            lbp = self._compute_lbp(gray)
            
            # Calculate histogram
            hist, _ = np.histogram(lbp.ravel(), bins=256, range=(0, 256))
            hist = hist.astype("float")
            hist /= (hist.sum() + 1e-7)
            
            # Calculate entropy (higher entropy = more texture = more likely real)
            entropy = -np.sum(hist * np.log2(hist + 1e-7))
            
            # Normalize entropy to [0, 1]
            max_entropy = np.log2(256)
            normalized_entropy = entropy / max_entropy
            
            return normalized_entropy
            
        except Exception as e:
            logger.error(f"Texture analysis error: {e}")
            return 0.5
    
    def _compute_lbp(self, image: np.ndarray, radius: int = 1, n_points: int = 8) -> np.ndarray:
        """Compute Local Binary Pattern"""
        h, w = image.shape
        lbp = np.zeros((h, w), dtype=np.uint8)
        
        for i in range(radius, h - radius):
            for j in range(radius, w - radius):
                center = image[i, j]
                code = 0
                
                for k in range(n_points):
                    angle = 2 * np.pi * k / n_points
                    x = int(round(i + radius * np.cos(angle)))
                    y = int(round(j + radius * np.sin(angle)))
                    
                    if 0 <= x < h and 0 <= y < w:
                        if image[x, y] >= center:
                            code |= (1 << k)
                
                lbp[i, j] = code
        
        return lbp
    
    def _analyze_color_space(self, face_roi: np.ndarray) -> float:
        """
        Analyze color distribution
        Real faces have specific color characteristics in different color spaces
        """
        try:
            # Convert to different color spaces
            hsv = cv2.cvtColor(face_roi, cv2.COLOR_BGR2HSV)
            ycrcb = cv2.cvtColor(face_roi, cv2.COLOR_BGR2YCrCb)
            
            # Analyze skin color in YCrCb space
            # Typical skin color ranges: Cr=[133-173], Cb=[77-127]
            cr = ycrcb[:, :, 1]
            cb = ycrcb[:, :, 2]
            
            # Calculate percentage of pixels in skin color range
            skin_mask = ((cr >= 133) & (cr <= 173) & (cb >= 77) & (cb <= 127))
            skin_percentage = np.sum(skin_mask) / skin_mask.size
            
            # Analyze color variance
            color_std = np.std(face_roi, axis=(0, 1))
            color_variance = np.mean(color_std) / 255.0
            
            # Combine metrics
            score = (skin_percentage * 0.6 + color_variance * 0.4)
            
            return min(1.0, score)
            
        except Exception as e:
            logger.error(f"Color space analysis error: {e}")
            return 0.5
    
    def _analyze_frequency(self, face_roi: np.ndarray) -> float:
        """
        Analyze frequency domain
        Printed photos have different frequency characteristics than real faces
        """
        try:
            gray = cv2.cvtColor(face_roi, cv2.COLOR_BGR2GRAY)
            
            # Apply FFT
            f_transform = np.fft.fft2(gray)
            f_shift = np.fft.fftshift(f_transform)
            magnitude = np.abs(f_shift)
            
            # Analyze high frequency components
            h, w = magnitude.shape
            center_h, center_w = h // 2, w // 2
            
            # Define high frequency region (outer 30%)
            mask = np.zeros((h, w), dtype=np.uint8)
            cv2.circle(mask, (center_w, center_h), int(min(h, w) * 0.35), 1, -1)
            cv2.circle(mask, (center_w, center_h), int(min(h, w) * 0.15), 0, -1)
            
            # Calculate high frequency energy
            high_freq_energy = np.sum(magnitude * mask)
            total_energy = np.sum(magnitude)
            
            high_freq_ratio = high_freq_energy / (total_energy + 1e-7)
            
            # Real faces typically have more high frequency content
            score = min(1.0, high_freq_ratio * 10)
            
            return score
            
        except Exception as e:
            logger.error(f"Frequency analysis error: {e}")
            return 0.5
    
    def _detect_moire_pattern(self, face_roi: np.ndarray) -> float:
        """
        Detect Moiré patterns (common in photos of screens/photos)
        """
        try:
            gray = cv2.cvtColor(face_roi, cv2.COLOR_BGR2GRAY)
            
            # Apply bandpass filter to detect periodic patterns
            # Moiré patterns appear as regular wave-like patterns
            
            # Compute gradient
            grad_x = cv2.Sobel(gray, cv2.CV_64F, 1, 0, ksize=3)
            grad_y = cv2.Sobel(gray, cv2.CV_64F, 0, 1, ksize=3)
            
            # Compute gradient magnitude
            grad_mag = np.sqrt(grad_x**2 + grad_y**2)
            
            # Analyze gradient variance
            # Moiré patterns have high periodic variance
            grad_std = np.std(grad_mag)
            grad_mean = np.mean(grad_mag)
            
            if grad_mean > 0:
                coefficient_of_variation = grad_std / grad_mean
            else:
                coefficient_of_variation = 0
            
            # Lower CV suggests less periodic patterns (more likely real)
            # Normalize and invert (higher score = more likely real)
            score = max(0.0, 1.0 - min(1.0, coefficient_of_variation / 2.0))
            
            return score
            
        except Exception as e:
            logger.error(f"Moiré pattern detection error: {e}")
            return 0.5
    
    def detect_eye_blink(self, frames: list) -> bool:
        """
        Detect eye blink across multiple frames
        This requires video input (multiple frames)
        
        Args:
            frames: List of consecutive frames
        
        Returns:
            True if blink detected, False otherwise
        """
        try:
            if len(frames) < 3:
                logger.warning("Not enough frames for blink detection")
                return False
            
            # Load eye cascade classifier
            eye_cascade = cv2.CascadeClassifier(
                cv2.data.haarcascades + 'haarcascade_eye.xml'
            )
            
            eye_states = []
            
            for frame in frames:
                gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
                eyes = eye_cascade.detectMultiScale(gray, 1.3, 5)
                
                # If eyes detected, mark as open (1), else closed (0)
                eye_states.append(1 if len(eyes) >= 2 else 0)
            
            # Check for blink pattern: open -> closed -> open
            # Pattern: [1, 1, 0, 1, 1] or similar
            for i in range(len(eye_states) - 2):
                if eye_states[i] == 1 and eye_states[i+1] == 0 and eye_states[i+2] == 1:
                    return True
            
            return False
            
        except Exception as e:
            logger.error(f"Eye blink detection error: {e}")
            return False
    
    def get_liveness_score_detailed(self, image: np.ndarray, face: Tuple[int, int, int, int]) -> dict:
        """
        Get detailed liveness scores for debugging
        
        Args:
            image: Input image
            face: Face bounding box
        
        Returns:
            Dictionary with detailed scores
        """
        try:
            x, y, w, h = face
            face_roi = image[y:y+h, x:x+w]
            
            if face_roi.size == 0:
                return {"error": "Empty face ROI"}
            
            texture_score = self._analyze_texture(face_roi)
            color_score = self._analyze_color_space(face_roi)
            frequency_score = self._analyze_frequency(face_roi)
            moire_score = self._detect_moire_pattern(face_roi)
            
            weights = [0.3, 0.25, 0.25, 0.2]
            scores = [texture_score, color_score, frequency_score, moire_score]
            final_score = sum(s * w for s, w in zip(scores, weights))
            
            return {
                "texture_score": round(texture_score, 3),
                "color_score": round(color_score, 3),
                "frequency_score": round(frequency_score, 3),
                "moire_score": round(moire_score, 3),
                "final_score": round(final_score, 3),
                "is_live": final_score > 0.5
            }
            
        except Exception as e:
            logger.error(f"Detailed liveness score error: {e}")
            return {"error": str(e)}