def identify_faces(image, feature_type, classifier_name):
""" Function detects and classifies faces in an image """
model_file = cfg.ssd_model
prototxt_file = cfg.prototxt_file
if np.shape(image)[0] >= np.shape(image)[1]:
#individual photo config
merge_overlap = 0.1
aspect_ratio_bounds = (0.4, 2)
min_confidence = 0.6
step_size = 1000
window_size = (2000, 2000)
else:
# group photo config
merge_overlap = 0.2
aspect_ratio_bounds = (0.5, 1.7)
min_confidence = 0.5
ws = int(round(np.shape(image)[0] / 6))
step_size = int(round(ws / 2))
window_size = (ws, ws)
face_bboxes = detect_faces(image, model_file, prototxt_file,
min_confidence, aspect_ratio_bounds,
merge_overlap, step_size, window_size)
for face_bbox in face_bboxes:
x1, y1, x2, y2 = face_bbox
face_img = image[y1:y2, x1:x2]
#skip empty bboxes
if any([x == 0 for x in np.shape(face_img)]):
continue
x = int(round(np.mean([x1, x2])))
y = int(round(np.mean([y1, y2])))
face_id = classify_face(face_img,
method=classifier_name,
feature_type=feature_type)
#u.imshow(face_img)
#remove unknowns and convert to 2-character ID
face_id = int(
face_id
) if 'unknown' not in face_id else cfg.unknown_face_return_value
this_result = np.array([[face_id, x, y]], dtype=np.int64)
#concat to results matrix if it exists, otherwise create it
if 'result_mat' not in locals():
result_mat = this_result
else:
result_mat = np.concatenate([result_mat, this_result], axis=0)
#if no faces found, return empty array
if len(face_bboxes) == 0 or 'result_mat' not in locals():
#if no faces detected or all face bboxes detected invalid (1 dim ==0)
result_mat = np.array([[]])
return result_mat
def recognize(img_vec, tags, mean, eig_vec):
face_classifier_name = 'haarcascade_frontalface_default.xml'
face_classifier = cv2.CascadeClassifier(face_classifier_name)
# Capture video
frame_first_flg = 1
cap = cv2.VideoCapture(0)
while True:
ret, frame = cap.read()
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
faces_vec = detect_faces(gray, face_classifier)
cv2.imshow("frame", gray)
key = cv2.waitKey(1)
if key & 0xFF == ord(' '):
if not faces_vec:
print("No faces detected. Try again.")
else:
# Weights of the dataset
pca_data = img_vec - mean
data_weights = np.transpose(
np.dot(eig_vec, np.transpose(pca_data)))
img_idx = 0
for face in faces_vec:
# Weights of test image
pca_data = face - mean
test_weights = []
for i in range(len(eig_vec)):
test_weights.append(
np.dot(eig_vec[i], np.transpose(pca_data)))
# Euclidean distance between weights
dist = []
for idx in range(len(data_weights)):
diff = np.linalg.norm(test_weights - data_weights[idx])
dist.append(diff)
min_dist = np.argmin(dist)
recognized_tag = tags[np.argmin(dist)]
img_idx += 1
print("Recognized person #{}: {}".format(
img_idx, recognized_tag))
if key == 27:
break
cap.release()
cv2.destroyAllWindows()
def correct_eyes_face(image, underline):
img0 = image.copy()
img = image
#Faces
faces = detect_faces(img, underline)
for (x, y, w, h) in faces:
face_area = img[y:y + h, x:x + w]
face_area0 = img0[y:y + h, x:x + w]
#Eyes in the face
eyes = detect_eyes(face_area0, underline)
correct_eyes(face_area, eyes)
#Underline borders
if underline:
for (x, y, w, h) in faces:
face_area = img[y:y + h, x:x + w]
cv2.rectangle(img, (x, y), (x + w, y + h), (255, 0, 0), 2)
for (ex, ey, ew, eh) in eyes:
cv2.rectangle(face_area, (ex, ey), (ex + ew, ey + eh),
(0, 255, 0), 2)
def get_similar_faces(img, max_results, query_dets):
result_lst = []
score_lst = []
if query_dets == " ":
# 1 - Detect face in query image (dlib or Retina face)
try:
detector = init_detector()
dets = detect_faces(img, detector)
except:
raise DetectorNotFoundError()
else:
dets = convert_bbox_string(query_dets)
if len(dets) == 0:
raise NoFaceInImageError()
# 2 - Align face
try:
fa = init_aligner()
faces = align_faces_sim_search(img, dets, fa)
face = faces[0]
except:
raise AlignerNotFoundError()
# 3 - Extract feature vector for query face
try:
face_embedding = embed_face(face)
except:
raise ModelNotFoundError()
# 4 - similarity search on faiss index
try:
index = load_index(PATH_INDEX_FILE, MODE)
lbs_index = load_labels(PATH_LABELS_FILE)
except:
raise IndexNotFoundError()
result_lst = similarity_search(index, lbs_index, [face_embedding],
max_results)
return result_lst
def add_people(args):
print("Parsing data...")
parser = add_parser()
args = parse_args(parser)
print("Reading data...")
if not os.path.isfile(args.new_data_file):
copyfile(args.old_data_file, args.new_data_file)
data = CsvData("image_data")
data.read(args.new_data_file)
print("Adding images to the dataset...")
face_classifier_name = 'haarcascade_frontalface_default.xml'
face_classifier = cv2.CascadeClassifier(face_classifier_name)
tag_prev = ''
for root, dir, files in os.walk(args.new_db_path):
for file in files:
tag = root[root.rfind('/') + 1:]
if tag != tag_prev:
tag_prev = tag
if tag in data.x:
print("{} already in the database. "
"Omitting...".format(tag))
continue
print("Adding {}...".format(tag))
img = cv2.imread(os.path.join(root, file), 0)
img = detect_faces(img, face_classifier, detect_one=True)
[img_w, img_h] = np.shape(img)
img_vec = np.reshape(img, img_w * img_h)
# Save tag and image vector to a csv file
df = pd.DataFrame({'tag': [tag], 'face': [img_vec.tolist()]})
with open(args.new_data_file, 'a') as f:
df.to_csv(f, header=False, index=False)
print("Done.")
def analyze_video(queue, video_file, settings):
"""
Finds faces' locations and encodings in the desired frames, and finds unique faces
:param queue: Multiprocessing queue
:param video_file: Path to input video file
:param settings: Settings dictionary
"""
# get video
video = cv2.VideoCapture(video_file) # input VideoCapture object
frame_count = int(video.get(cv2.CAP_PROP_FRAME_COUNT)) # number of frames in input video
# initialize models
net = models.detection_dnn_model() # detection model
face_encoder = models.face_encoder() # face recognition encoding model
pose_predictor = models.pose_predictor() # pose predictor model for finding face landmarks
# get settings
track_period = settings["track_period"] # track period from settings
resamples = settings["resamples"] # number of resamples from settings
tolerance = settings["tolerance"] # face matching tolerance from settings
# initialize lists
face_locations = [] # list of face locations for each frame
face_encodings = [] # list of face encodings for each frame
unique_frames = [] # list of frame images including unique faces
unique_face_locations = [] # list of unique faces' locations
unique_encodings = [] # list of unique faces' encodings
for i in range(frame_count):
ret, img = video.read() # ret indicates if frame was read correctly, img is last read frame
if i % track_period == 0: # frame for detection
rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # frame in rgb format
face_locations.append(detection.detect_faces(img, net))
face_encodings.append(recognition.encode_faces(rgb, face_locations[-1], pose_predictor, face_encoder, resamples))
new_indices, new_encodings = recognition.exclude_faces(face_encodings[-1], np.array(unique_encodings), tolerance) # indices and encodings of new faces found
for k in range(len(new_indices)): # for each new face found
unique_frames.append(rgb)
unique_face_locations.append(face_locations[-1][k])
unique_encodings.append(new_encodings[k])
send_data(queue, [face_locations, face_encodings, unique_frames, unique_face_locations, unique_encodings])
print("Start Reading...")
while True:
_, img = video.read()
img = cv2.transpose(img)
img = cv2.flip(img, 1)
height, width, channel = img.shape
# matrix = cv2.getRotationMatrix2D((width / 2, height / 2), 270, 1)
# frame = cv2.warpAffine(frame, matrix, (width, height))
tic = time.time()
resize_img = cv2.resize(img, (0, 0), fx=resize_rate, fy=resize_rate)
if resize_img.ndim == 2:
resize_img = facenet.to_rgb(resize_img)
resize_img = resize_img[:, :, 0:3]
bounding_boxes = detection.detect_faces(resize_img, img.shape)
if bounding_boxes.shape[0] > 0:
match_names, p = recognition.recognize_faces(img, bounding_boxes)
else:
bounding_boxes = match_names = p = []
toc = time.time() - tic
img = utils.mosaic(img, bounding_boxes, match_names, 6)
img = utils.draw_box(img, bounding_boxes, match_names, p)
cv2.imshow("origin", img)
cv2.waitKey(0)