def detect_faces(self, image, conf_threshold: float = 0.7) -> List[List[int]]:
"""Performs facial detection on an image. Uses MTCNN.
Args:
image (numpy array):
conf_threshold (float, optional): Threshold confidence to consider
Raises:
InvalidImage: When the image is either None or
with wrong number of channels.
Returns:
List[List[int]]: List of bounding box coordinates
"""
if not is_valid_img(image):
raise InvalidImage
# Do a forward propagation with the blob created from input img
detections = self.face_detector.detect_faces(image)
# Bounding box coordinates of faces in image
bboxes = []
for _, detection in enumerate(detections):
conf = detection["confidence"]
if conf >= conf_threshold:
x, y, w, h = detection["box"]
x1, y1, x2, y2 = x, y, x + w, y + h
# Trim forehead area to match dlib style facial ROI
if self.crop_forehead:
y1 = y1 + int(h * self.shrink_ratio)
bboxes.append([x1, y1, x2, y2])
return bboxes
def register_face(self,
image=None,
name: str = None,
bbox: List[int] = None):
"""Method to register a face via the facial encoding.
Siamese neural network is used to generate 128 numbers
for a given facial region. These encodings can be used to identify a
facial ROI for identification later.
Args:
image (numpy array, optional): Defaults to None.
name (str, optional): Name to associate with the face. Defaults to None.
bbox (List[int], optional): Facial ROI bounding box. Defaults to None.
Raises:
NoNameProvided:
NoFaceDetected:
Returns:
Dict: Facial encodings along with an unique identifier and name
"""
if not is_valid_img(image) or name is None:
raise NoNameProvided if name is None else InvalidImage
image = image.copy()
face_encoding = None
try:
if bbox is None:
bboxes = self.face_detector.detect_faces(image=image)
if len(bboxes) == 0:
raise NoFaceDetected
bbox = bboxes[0]
face_encoding = self.get_facial_fingerprint(image, bbox)
# Convert the numpy array to normal python float list
# to make json serialization simpler
facial_data = {
"id": str(uuid.uuid4()),
"encoding": tuple(face_encoding.tolist()),
"name": name,
}
# save the encoding with the name
self.save_facial_data(facial_data)
logger.info("Face registered with name: {}".format(name))
except Exception as exc:
raise exc
return facial_data
def convert_to_rgb(image):
"""Converts an image to RGB format.
Args:
image (numpy array): [description]
Raises:
InvalidImage: [description]
Returns:
[type]: [description]
"""
if not is_valid_img(image):
raise InvalidImage
return cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
def detect_faces(self,
image,
conf_threshold: float = 0.7) -> List[List[int]]:
"""Performs facial detection on an image. Uses OpenCV DNN based face detector.
Args:
image (numpy array):
conf_threshold (float, optional): Threshold confidence to consider
Raises:
InvalidImage: When the image is either None or
with wrong number of channels.
Returns:
List[List[int]]: List of bounding box coordinates
"""
if not is_valid_img(image):
raise InvalidImage
# To prevent modification of orig img
image = image.copy()
height, width = image.shape[:2]
# Do a forward propagation with the blob created from input img
detections = self.model_inference(image)
# Bounding box coordinates of faces in image
bboxes = []
for idx in range(detections.shape[2]):
conf = detections[0, 0, idx, 2]
if conf >= conf_threshold:
# Scale the bbox coordinates to suit image
x1 = int(detections[0, 0, idx, 3] * width)
y1 = int(detections[0, 0, idx, 4] * height)
x2 = int(detections[0, 0, idx, 5] * width)
y2 = int(detections[0, 0, idx, 6] * height)
if self.crop_forehead:
y1 = y1 + int(height * self.shrink_ratio)
# openCv detector can give a lot of false bboxes
# when the image is a zoomed in face / cropped face
# This will get rid of atleast few, still there can be other
# wrong detections present!
if self.is_valid_bbox([x1, y1, x2, y2], height, width):
bboxes.append([x1, y1, x2, y2])
return bboxes
def detect_faces(self, image, num_upscaling: int = 1) -> List[List[int]]:
"""Performs facial detection on an image. Works best with
RGB image. Uses a dlib based detector either HOG or MMOD.
Args:
image (numpy array):
num_upscaling (int, optional): Number of times to upscale
while detecting faces. Defaults to 1.
Raises:
InvalidImage: When the image is either None or
with wrong number of channels.
Returns:
List[List[int]]: List of bounding box coordinates
"""
if not is_valid_img(image):
raise InvalidImage
return [
self.dlib_rectangle_to_list(bbox)
for bbox in self.face_detector(image, num_upscaling)
]