Spaces:
Sleeping
Sleeping
| """ | |
| face_utils.py — Face detection and crop using OpenCV. | |
| Detects the largest frontal face in an image and returns a padded crop. | |
| Returns (crop, face_found) where face_found=False means no person in frame. | |
| NO edge mask is applied — raw pixel data is preserved so MobileNetV2 | |
| sees the same features during training and prediction. | |
| """ | |
| import cv2 | |
| import numpy as np | |
| # OpenCV's built-in Haar cascades — no download needed | |
| _frontal_cascade = cv2.CascadeClassifier( | |
| cv2.data.haarcascades + 'haarcascade_frontalface_default.xml' | |
| ) | |
| _alt_cascade = cv2.CascadeClassifier( | |
| cv2.data.haarcascades + 'haarcascade_frontalface_alt2.xml' | |
| ) | |
| def detect_and_crop_face(img_rgb: np.ndarray, padding: float = 0.25): | |
| """ | |
| Detect the largest face in img_rgb (uint8 RGB HxWx3) and return a padded crop. | |
| Args: | |
| img_rgb: uint8 numpy array in RGB colour space, shape (H, W, 3) | |
| padding: fractional padding added around the detected face box (0.25 = 25%) | |
| Returns: | |
| (cropped_rgb, face_found) | |
| - cropped_rgb: the face crop, or the original image if no face detected | |
| - face_found: bool — False means no person in frame → caller returns Unknown | |
| """ | |
| if img_rgb is None or img_rgb.size == 0: | |
| return img_rgb, False | |
| gray = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2GRAY) | |
| gray = cv2.equalizeHist(gray) # improves detection in poor lighting | |
| faces = [] | |
| # Try primary frontal cascade | |
| detected = _frontal_cascade.detectMultiScale( | |
| gray, scaleFactor=1.1, minNeighbors=4, minSize=(40, 40) | |
| ) | |
| if len(detected) > 0: | |
| faces = detected | |
| # If no face found, try the alt cascade (detects tilted / partially visible faces) | |
| if len(faces) == 0: | |
| detected2 = _alt_cascade.detectMultiScale( | |
| gray, scaleFactor=1.1, minNeighbors=4, minSize=(40, 40) | |
| ) | |
| if len(detected2) > 0: | |
| faces = detected2 | |
| if len(faces) == 0: | |
| # No face detected at all → caller should return Unknown | |
| return img_rgb, False | |
| # Pick the largest detected face | |
| x, y, w, h = max(faces, key=lambda f: f[2] * f[3]) | |
| # Apply padding around the face box | |
| pad_x = int(w * padding) | |
| pad_y = int(h * padding) | |
| H, W = img_rgb.shape[:2] | |
| x1 = max(0, x - pad_x) | |
| y1 = max(0, y - pad_y) | |
| x2 = min(W, x + w + pad_x) | |
| y2 = min(H, y + h + pad_y) | |
| face_crop = img_rgb[y1:y2, x1:x2] | |
| if face_crop.size == 0: | |
| return img_rgb, False | |
| return face_crop, True | |