def calculate_euclidean_distance(ref_frame, frame):
"""
128차원 embedding vector를 구하고 유클리드 거리 계산.
- input
reference_clip: VideoFileClip
clip: VideoFileClip
- output
euclidean distance between two vector, clip의 마지막 frame return
"""
# Get image as an numpy array
#ref_frame = reference_clip.get_frame(reference_clip.reader.nframes/reference_clip.fps)
#frame = clip.get_frame(clip.reader.nframes/clip.fps)
# detection
ref_frame_detected = extract_face_from_frame(ref_frame)
frame_detected = extract_face_from_frame(frame)
if (len(ref_frame_detected) == 0 or len(frame_detected) == 0): # 하나라도 얼굴이 감지되지 않는 경우
return 100
# feature extraction (embedding)
ref_frame_detected = ref_frame_detected.reshape((1,)+ref_frame_detected.shape) # 모델 인풋 차원에 맞게 수정
frame_detected = frame_detected.reshape((1,)+frame_detected.shape)
embed_model = FaceNet()
ref_frame_embed = embed_model.embeddings(ref_frame_detected)
frame_embed = embed_model.embeddings(frame_detected)
diff = ref_frame_embed - frame_embed
euclidean_dist = np.sqrt(np.sum(np.multiply(diff, diff)))
return euclidean_dist
def give_embeddings(image_links):
print("line1")
detector = MTCNN()
print("line2")
model = FaceNet()
print("line3")
embeddings_list = []
for image in image_links:
print("line4")
flag, faces_array = extract_faces(image, detector)
print("line5")
if (flag):
print("line6")
for face_pixels in faces_array:
print("line15")
face_pixels = face_pixels.astype('float32')
print("line16")
face_pixels = face_pixels.reshape(1, 160, 160, 3)
print("line17")
yhat = model.embeddings(face_pixels)
print("line18")
embeddings_list.append(yhat[0])
print("line19")
print("line20")
return embeddings_list
def extract_faceID(filename, id):
print("BAT DAU LAY THONG TIN FACE ID")
img = cv2.imread(filename)
embedder = FaceNet()
detector = MTCNN()
data = []
faces = detector.detect_faces(img)
for face in faces:
st = (face['box'][0], face['box'][1])
en = (face['box'][0] + face['box'][2], face['box'][1] + face['box'][3])
data.append(img[st[1]:en[1], st[0]:en[0]])
break
vt = embedder.embeddings(data)[0]
# save to database
sql = SQL_Server()
query = "INSERT INTO dbo.FaceID VALUES (" + str(id)
for i in vt:
query += " , " + str(i)
query += ")"
# print(query)
sql.insert(query)
print("ADD FACE ID INTO CSDL THANH CÔNG")
def _main():
# img_paths = generate_input_paths(dir_path, 138)
# for ip in img_paths:
# faces = process_image(ip)
with open(sys.argv[1], 'rb') as f:
(le, model) = pickle.load(f, encoding='latin1')
embedder = FaceNet()
directory = sys.argv[2]
files = sorted(os.listdir(directory))
for i, filename in enumerate(files):
path = directory + filename
boxes = process_image(path)
img = np.asarray(Image.open(path).convert('RGB'))
result = []
for box in boxes:
x1, y1, width, height = box
x1, y1 = abs(x1), abs(y1)
x2, y2 = x1 + width, y1 + height
face = img[y1:y2, x1:x2]
face = Image.fromarray(face).resize((IMAGE_SIZE, IMAGE_SIZE))
face = np.asarray(face)
rep = embedder.embeddings([face])
pred = model.predict_proba(rep).ravel()
maxI = np.argmax(pred)
confidence = pred[maxI]
person = le.inverse_transform([maxI])[0]
result.append('{} ({:.2f})'.format(person, confidence))
print(i, ', '.join(result))
def trackimage():
embedder = FaceNet()
b = []
cap = cv2.VideoCapture(0)
while True:
ret, frame = cap.read()
img1 = extract_face(frame)
plt.imshow(frame)
img1 = np.expand_dims(img1, axis=0)
if (img1.any()):
emb = embedder.embeddings(img1)
emb = np.transpose(emb)
min_dist = 100
for key, value in dictq.items():
dist = np.linalg.norm(emb - value)
b.append(dist)
if dist < min_dist:
min_dist = dist
identity = key
print(identity)
if min_dist < 1.0:
cv2.putText(frame, "Face : " + identity, (100, 100),
cv2.FONT_HERSHEY_PLAIN, 1.5, (0, 0, 255), 2)
unknown_yes_or_no = 'no'
else:
cv2.putText(frame, 'no match', (100, 100),
cv2.FONT_HERSHEY_PLAIN, 1.5, (0, 0, 255), 2)
unknown_yes_or_no = 'yes'
cv2.imshow('face', frame)
if cv2.waitKey(1) & 0xFF == 27:
break
cap.release()
cv2.destroyAllWindows()
import pyttsx3
engine = pyttsx3.init()
if (unknown_yes_or_no == "yes"):
engine.say("Good morning sorry we couldn't recognise you")
else:
str1 = "good morning " + identity + " your attendance has been recorded"
engine.say(str1)
import mysql.connector
mydb = mysql.connector.connect(host="localhost",
user="******",
passwd="albertbolt23",
database="faceproject",
auth_plugin='mysql_native_password')
mycursor = mydb.cursor()
from datetime import date
today1 = date.today()
sql = "INSERT INTO attendance values('%s','%s','morning')" % (
identity, str(today1))
mycursor.execute(sql)
mydb.commit()
engine.runAndWait()
def save_face_encoding(src_path):
facenet = FaceNet()
img_files = glob.glob('{0}/**/*.jpg'.format(src_path), recursive=True)
for img_file in tqdm(img_files):
em = facenet.embeddings([cv2.imread(img_file)[:, :, ::-1]])
np.save(img_file.replace('jpg', 'npy'), em[0])
def create_pickle(training_path, testing_path, training_destination,
testing_destination):
embedder = FaceNet()
# images is a list of images, each as an
# np.ndarray of shape (H, W, 3).
for paths, destination in [(training_paths, training_destination),
(testing_paths, testing_destination)]:
print("Getting images...")
images = get_images_from_filepaths(paths)
print("Getting embeddings...")
try:
embeddings = embedder.embeddings(images, verbose=1)
except:
embeddings = embedder.embeddings(images)
print("Saving embeddings...")
save_imgs(paths, embeddings, destination)
class FaceNetWrap:
def __init__(self, callback=None):
self.file_list = os.listdir('./dataset/train')
self.faces = []
self.names = [] # faces - names相同下标位置一一对应
for file in self.file_list:
self.faces.append(convert_gray_to_bgr_use_path(os.path.join('./dataset/train', file)))
self.names.append(file.split('.')[0].split('-')[1])
self.embedder = FaceNet()
self.embeddings = self.embedder.embeddings(self.faces)
if callback:
callback()
def get_embedding(self, img):
embeddings_test = self.embedder.embeddings([img])
return embeddings_test[0]
# 返回匹配人名和距离、照片
def get_best_fit(self, embedding):
dis = [get_L2_norm_squared(em, embedding) for em in self.embeddings]
idx = np.argmin(dis)
img = cv2.imread(os.path.join('./dataset/train', self.file_list[idx]), cv2.IMREAD_GRAYSCALE)
name = self.names[idx]
return name, dis[idx], img
def createDatabase():
database = pd.DataFrame({'folder_id':[],'photo_id':[],'face_id':[],'embedding':[]},dtype="float32")
#traverse through all the images
images_path = "/home/darealappu/Desktop/CDAC/DR-GAN tensorflow/vggface2_test/test"
index=0
time_to_return = False
embedder = FaceNet()
for folder in os.listdir(images_path):
folder_id = int(folder[1:])
for image in os.listdir(os.path.join(images_path,folder)):
if time_to_return == True:
return database
photo_id = int(image[0:4])
face_id = int(image[5:7])
img = cv2.imread(os.path.join(images_path,folder,image)).astype(np.float32)
img.resize((1,img.shape[0],img.shape[1],img.shape[2]))
em = embedder.embeddings(img)
database=database.append({
'folder_id':folder_id,
'photo_id':photo_id,
'face_id':face_id,
'embedding':em},ignore_index=True)
index+=1
print(index-1)
if index%10==0:
if(index==5000):
time_to_return = True
break;
return database
# 'sl08afxcx4_115.515400-119.986533','bLEddi92aFI_171.480000-177.400000']
faces2_folder = '../../../../final_data/separated_data/faces2'
paths = []
for (dirpath, dirnames, filenames) in os.walk(faces2_folder):
paths.extend(dirnames)
break
count = 0
for path in paths:
print("{0}/{1}".format(count, len(paths)))
count += 1
a = np.zeros((75, 512))
for i in range(75):
image = cv2.imread('../../../../final_data/separated_data/faces2/'+path+'/face'+str(i)+'.jpg')
# gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
image = np.array(image)
print(image.shape)
tmp = []
tmp.append(image)
a[i,:] = embedder.embeddings(tmp)
try:
os.makedirs('../../../../final_data/separated_data/nparray')
except FileExistsError:
pass
np.save('../../../../final_data/separated_data/nparray/'+path+'_facenet.npy',a)
# with open('../../../download/separated_data/faces2/'+path+'/facenet_out.csv', mode='a') as f:
# facenet_writer = csv.writer(f, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL)
# facenet_writer.writerow([path,str(i),embeddings])
import pickle
import matplotlib.pyplot as plt
import numpy as np
import mtcnn
import tensorflow as tf
from PIL import Image
physical_devices = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)
from sklearn.preprocessing import Normalizer
from keras_facenet import FaceNet
embedder = FaceNet()
normalizer = Normalizer(norm='l2')
from detect_face import face_extract
class_label = [
'Ben Affleck', 'Elton John', 'Jerry Seinfeld', 'Madonna', 'Mindy Kaling'
]
with open('model.pkl', 'rb') as f:
clf = pickle.load(f)
file = 'dataset/val/madonna/httpassetsrollingstonecomassetsarticlemadonnadavidbowiechangedthecourseofmylifeforeversmallsquarexmadonnabowiejpg.jpg'
img = plt.imread(file)
face = face_extract(file, (160, 160))
face_array = np.asarray(face)
face_array = np.expand_dims(face_array, axis=0)
emded = embedder.embeddings(face_array)
emded = normalizer.fit_transform(emded)
pred = clf.predict(emded)
label = class_label[pred.argmax()]
print(label)
import tensorflow as tf
from PIL import Image
physical_devices = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], True)
from sklearn.metrics import accuracy_score
from sklearn.preprocessing import LabelEncoder
from sklearn.preprocessing import Normalizer
from sklearn.svm import SVC
import pickle
from keras_facenet import FaceNet
embedder = FaceNet()
with open('data.pkl', 'rb') as f:
x_train, y_train, x_val, y_val = pickle.load(f)
emb_train_x = embedder.embeddings(x_train)
emb_val_x = embedder.embeddings(x_val)
normalizer = Normalizer(norm='l2')
x_train = normalizer.transform(emb_train_x)
x_val = normalizer.transform(emb_val_x)
encoder = LabelEncoder()
encoder.fit(y_train)
y_train = encoder.transform(y_train)
y_val = encoder.transform(y_val)
model = SVC(kernel='linear', probability=True)
model.fit(x_train, y_train)
# predict
prediction = model.predict(x_val)
def evaluate_id_lfw(dataset_path, cleanir, dsize=(64, 64)):
"""Evaluates CLEANIR model by using LFW dataset
Arguments:
dataset_path {str} -- lfw dataset path
cleanir {Cleanir} -- Cleanir instance
Keyword Arguments:
dsize {tuple} -- size of cropped face (default: {(64, 64)})
Returns:
dict -- evaluation results
"""
print('Reading LFW dataset pair information..')
matched_list, unmatched_list = read_lfw_pair_info(dataset_path)
print('Loading FaceNet..')
facenet = FaceNet()
o_dists = []
m0_dists = []
m30_dists = []
m60_dists = []
m90_dists = []
m120_dists = []
m150_dists = []
m180_dists = []
print('Loading and modifying LFW dataset images..')
for face1_path, face2_path in tqdm(matched_list):
face1 = crop_face_from_file(face1_path, dsize)
face2 = crop_face_from_file(face2_path, dsize)
deid = cleanir.get_deid_single_axis_func(face1)
ems = facenet.embeddings([
face1, face2,
deid(180),
deid(90),
deid(0),
deid(30),
deid(60),
deid(120),
deid(150)
])
o_dists += [facenet.compute_distance(ems[0], ems[1])]
m180_dists += [facenet.compute_distance(ems[2], ems[1])]
m90_dists += [facenet.compute_distance(ems[3], ems[1])]
m0_dists += [facenet.compute_distance(ems[4], ems[1])]
m30_dists += [facenet.compute_distance(ems[5], ems[1])]
m60_dists += [facenet.compute_distance(ems[6], ems[1])]
m120_dists += [facenet.compute_distance(ems[7], ems[1])]
m150_dists += [facenet.compute_distance(ems[8], ems[1])]
results = {
'threshold': [],
'original': [],
'0': [],
'30': [],
'60': [],
'90': [],
'120': [],
'150': [],
'180': []
}
print('Thresholding..')
for threshold in tqdm(np.arange(0.1, 2.0, 0.1)):
results['threshold'].append(threshold)
results['original'].append(
np.sum(np.array(o_dists) < threshold) / len(o_dists))
results['180'].append(
np.sum(np.array(m180_dists) < threshold) / len(m180_dists))
results['90'].append(
np.sum(np.array(m90_dists) < threshold) / len(m90_dists))
results['0'].append(
np.sum(np.array(m0_dists) < threshold) / len(m0_dists))
results['30'].append(
np.sum(np.array(m30_dists) < threshold) / len(m30_dists))
results['60'].append(
np.sum(np.array(m60_dists) < threshold) / len(m60_dists))
results['120'].append(
np.sum(np.array(m120_dists) < threshold) / len(m120_dists))
results['150'].append(
np.sum(np.array(m150_dists) < threshold) / len(m150_dists))
return results
class FaceRecognition(object):
"""
Face Recognition object class
"""
def __init__(self):
"""
Initialize Face Recognition model
"""
# GRAPH
self.graph = tf.get_default_graph()
# Load Face Detector
self.face_detector = MTCNN()
# Load FaceNet
self.facenet = FaceNet()
# Euclidean Classifier
self.clf = None
def predict(self, path, threshold=None):
"""
Find faces and recognize them, return predicted people into image
:param path: Source image path
:param threshold: cutoff threshold
:return: Return predictions and images with rectangles drawn
"""
if not self.clf:
raise RuntimeError("No classifier found. Please load classifier")
start_at = time.time()
bounding_boxes = []
image = cv2.imread(path)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = image.astype(np.uint8)
for person, confidence, box in self.__predict__(image,
threshold=threshold):
# Draw rectangle with person name
cv2.rectangle(image, (box[0], box[1]), (box[2], box[3]),
(0, 255, 0), 2)
cv2.putText(image, person, (box[0], box[1] - 10),
cv2.FONT_HERSHEY_SIMPLEX, 0.4, (0, 255, 0))
bounding_boxes.append({
"person": person,
"confidence": confidence,
"box": box,
})
# encode frame
_, buffer = cv2.imencode('.jpg', image)
return {
"frame": base64.b64encode(buffer).decode('ascii'),
"elapsed_time": (time.time() - start_at),
"predictions": bounding_boxes
}
def __predict__(self, image, threshold=None):
"""
Extract face and perform evaluation
:param image: Source image
:param threshold: decision threshold
:return: yield (person_id, person, confidence, box)
"""
# Resize Image
for encoding, face, box in self.face_encoding(image):
# Check face size
if (box[2] - box[0]) < config.MIN_FACE_SIZE[0] or \
(box[3] - box[1]) < config.MIN_FACE_SIZE[1]:
yield (config.UNKNOWN_LABEL, 0.0, box)
else:
results = self.clf.predict(encoding)
person, confidence = results["person"], results["confidence"]
if threshold and confidence < threshold:
person = config.UNKNOWN_LABEL
yield (person, confidence, box)
def face_detection(self, image):
"""
Face detection from source image
:param image: Source image
:return: extracted face and bounding box
"""
image_to_detect = image.copy()
# detect faces in the image
for face_attributes in self.face_detector.detect_faces(
image_to_detect):
if face_attributes["confidence"] > config.FACE_CONFIDENCE:
# extract the bounding box
x1, y1, w, h = [
max(point, 0) for point in face_attributes["box"]
]
x2, y2 = x1 + w, y1 + h
face = image[y1:y2, x1:x2]
# Align face
face = FaceRecognition.align_face(face_attributes, face.copy())
yield (cv2.resize(face, config.FACE_SIZE), (x1, y1, x2, y2))
def face_encoding(self, source_image):
"""
Extract face encodings from image
:param source_image: Source image
:return: 512 encoding, face and bounding box
"""
for face, box in self.face_detection(source_image):
with self.graph.as_default():
# Face encoding
encoding = self.facenet.embeddings(np.expand_dims(face,
axis=0))[0]
yield (encoding, face, box)
@staticmethod
def align_face(face_attribute, image):
if not face_attribute:
return image
# Get left and right eyes
left_eye = face_attribute["keypoints"]["left_eye"]
right_eye = face_attribute["keypoints"]["right_eye"]
# Get distance between eyes
d = math.sqrt(
math.pow(right_eye[0] - left_eye[0], 2) +
math.pow(right_eye[1] - left_eye[1], 2))
a = left_eye[1] - right_eye[1]
# get alpha degree
alpha = (math.asin(a / d) * 180.0) / math.pi
return imutils.rotate(image, -alpha)
def load(self, path):
"""
Load classifier from pickle file
:param path: path
"""
clf = EuclideanClassifier()
clf.load(path)
self.clf = clf
def save(self, path):
"""
Save classifier as pickle file
:param path: path
"""
self.clf.save(path)
def fit(self, folder):
"""
Fit classifier from directory.
Directory must have this structure:
Person 1:
file.jpg
....
file.jpg
Person 2:
file.jpg
...
file.jpg
...
:param folder: root folder path
"""
# Initialize classifier
clf = EuclideanClassifier()
# Load all files
files = []
for ext in config.ALLOWED_IMAGE_TYPES:
files.extend(
glob.glob(os.path.join(folder, "*", ext), recursive=True))
for path in tqdm.tqdm(files):
# Load image
image = cv2.imread(path)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# Get person name by folder
person = os.path.split(os.path.split(path)[0])[1]
# Get encoding
for encoding, face, box in self.face_encoding(image):
# Add to classifier
clf.fit([encoding], [person])
self.clf = clf
def fit_from_dataframe(self, df, person_col="person", path_col="path"):
"""
Fit classifier from dataframe.
:param df: Pandas dataframe
:param person_col: Dataframe column with person id
:param path_col: Dataframe column with image path
"""
# Initialize classifier
clf = EuclideanClassifier()
for index, row in tqdm.tqdm(df.iterrows(), total=df.shape[0]):
# Load image
image = cv2.imread(row[path_col])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# Get person name by folder
person = row[person_col]
# Get encoding
for encoding, face, box in self.face_encoding(image):
# Add to classifier
clf.fit([encoding], [person])
self.clf = clf
from keras_facenet import FaceNet
from utils import load_data
from scipy.spatial.distance import euclidean
from sklearn.preprocessing import normalize
import numpy as np
from tqdm import tqdm
train_images, train_labels, test_images, test_labels, train_pairs, submission_pairs = load_data(
)
embedder = FaceNet()
embeddings = embedder.embeddings(test_images)
embeddings = normalize(embeddings)
print(embeddings.shape)
test_labels_indices = {l: i for i, l in enumerate(test_labels)}
distances = []
for pair in tqdm(submission_pairs):
p1, p2 = pair
i1 = test_labels_indices[p1]
i2 = test_labels_indices[p2]
e1 = embeddings[i1]
e2 = embeddings[i2]
dist = euclidean(e1, e2)
distances.append(dist)
# convert distances to probability distribution
distances = np.asarray(distances)
import numpy as np
import sys
from keras.models import load_model
from keras_facenet import FaceNet
BATCH_SIZE = 1000
embedder = FaceNet()
if __name__ == '__main__':
data = np.load(sys.argv[1])
X, y = data['arr_0'], data['arr_1']
print('Loaded X: {}, y: {}'.format(X.shape, y.shape))
embeddings = []
for i in range(0, X.shape[0], BATCH_SIZE):
batch = X[i:i + BATCH_SIZE]
print(batch.shape)
batch_emb = np.asarray(embedder.embeddings(batch))
embeddings.extend(batch_emb)
np.savez_compressed(sys.argv[2], embeddings, y)
header=None)
y_test = y_test.values
return df, y_test
'''
Loading the data set
'''
X_train_orig, y_train = load_train_data(24)
X_test_orig, y_test = load_test_data(7)
print('X_Train data shape=', X_train_orig.shape)
print('X_Test data shape=', X_test_orig.shape)
print('y_Train data shape=', y_train.shape)
print('y_Test data shape=', y_test.shape)
X_train = embedder.embeddings(X_train_orig)
X_test = embedder.embeddings(X_test_orig)
print('Train embed shape=', X_train.shape)
print('Test embed shape=', X_test.shape)
from sklearn.svm import SVC
from sklearn.linear_model import LogisticRegression
model = LogisticRegression()
model.fit(X_train, y_train)
#Predicted faces
model.predict(X_test)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
results = detector.detect_faces(img)
if (results):
x1, y1, w, h = results[0]['box']
x1, y1 = abs(x1), abs(y1)
x2, y2 = x1 + w, y1 + h
# extract the face
face = img[y1:y2, x1:x2]
# resize pixels to the model size
image = cv2.resize(face, required_size, cv2.INTER_AREA)
return image
x = []
for i in glob.glob(r'C:\Users\albertbolt\Downloads\face_dataset\*\*.jpeg'):
img = cv2.imread(i)
img = extract_face(img)
x.append(img)
x = np.stack(x)
embedder = FaceNet()
embeddings = embedder.embeddings(x)
a = []
for root, dirp, file in os.walk(r'C:\Users\albertbolt\Downloads\face_dataset'):
a.append(dirp)
dictq = {}
for i in range(len(a[0])):
dictq[a[0][i]] = embeddings[i]
for key, value in dictq.items():
value.shape = [512, 1]
print(dictq)
class FaceRegModel(object):
def __init__(self):
self.model = MTCNN()
self.embedder = FaceNet()
@staticmethod
def load_from_file(f):
img_str = f.read()
nparr = np.fromstring(img_str, np.uint8)
img_np = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
cv2.cvtColor(img_np, cv2.COLOR_BGR2RGB)
return img_np
def encode_face(self, img):
location = self.locate_faces(img)[:1]
face_img = self.extract_faces(img, location)
return self.embedder.embeddings(face_img)[0]
def encode_faces(self, img, locations):
return self.embedder.embeddings(self.extract_faces(img, locations))
def extract_faces(self, img, locations):
res_imgs = []
for (y1, x1, y2, x2) in locations:
res_imgs.append(img[y1:y2, x2:x1])
return res_imgs
def locate_faces(self, img):
faces = self.model.detect_faces(img)
faces = [x['box'] for x in faces]
faces = [(max(y, 0), min(x + w,
img.shape[1]), min(y + h,
img.shape[0]), max(x, 0))
for x, y, w, h in faces]
return np.array(faces)
@staticmethod
def face_distance(face1, face2):
return np.linalg.norm(face1 - face2, axis=1)
@staticmethod
def compare_faces(face1, face2, tolerance=0.9):
return list(FaceRegModel.face_distance(face1, face2) <= tolerance)
def process_images(self, images, scale=0.5):
"""Get face locations and encodings from given images."""
face_information = {}
for (img_name, img) in images.items():
# Scale down for faster processing
small_img = cv2.resize(img, (0, 0), fx=scale, fy=scale)
# Face detection
face_locations = self.locate_faces(small_img)
if len(face_locations) == 0:
# Ignoring empty
face_encodings = []
else:
# Face extraction
face_encodings = self.encode_faces(small_img, face_locations)
face_information[img_name] = {
'locations': face_locations,
'encodings': face_encodings
}
return face_information
def find_images_with_person(self, face_encoding, face_information):
images_with_person = []
for img_name, face_info in face_information.items():
face_info_encodings = face_info
if len(face_info_encodings) == 0:
continue
matches = self.compare_faces(face_encoding, face_info_encodings)
if any(matches):
images_with_person.append(img_name)
return images_with_person
"static/img/din.jpg", "static/img/ani.jpg", "static/img/sidh.jpg"
]
images = []
for i in tqdm(images_path):
print(path + i)
img = cv2.imread(path + i)
img = cv2.resize(img, (480, 480))
img = numpy.asarray(img)
images.append(img)
images = np.array(images)
print(images.shape)
# images is a list of images, each as an
# np.ndarray of shape (H, W, 3).
embeddings = embedder.embeddings(images)
import math
print(embeddings.shape)
# print(embeddings[0][:50])
temp1 = 0
temp2 = 0
temp3 = 0
temp4 = 0
temp5 = 0
for i in range(len(embeddings[0])):
temp1 += ((embeddings[0][i]) - (embeddings[1][i])) #--same
temp2 += ((embeddings[0][i]) - (embeddings[2][i])) #diff
temp3 += ((embeddings[0][i]) - (embeddings[3][i])) #diff
temp4 += ((embeddings[0][i]) - (embeddings[4][i])) #diff
temp5 += ((embeddings[0][i]) - (embeddings[5][i])) #diff
if __name__ == '__main__':
with open(sys.argv[1], 'rb') as f:
(le, model) = pickle.load(f, encoding='latin1')
embedder = FaceNet()
directory = sys.argv[2]
files = sorted(os.listdir(directory))
for i, filename in enumerate(files):
path = directory + filename
faces, d_conf, d_loc, img = detect_faces(path, min_conf)
d = Draw(img)
results = []
for j, face in enumerate(faces):
rep = embedder.embeddings([face])
pred = model.predict_proba(rep).ravel()
maxI = np.argmax(pred)
confidence = pred[maxI]
person = le.inverse_transform([maxI])[0]
result = '{} ({:.2f})'.format(person, confidence)
d.rectangle(d_loc[j], outline='green')
d.text(d_loc[j][0],
'{}\n detect: {:.2f}'.format(result, d_conf[j]))
results.append(result)
print(i, ', '.join(results))
if sys.argv[3]:
img.save('{}/{}'.format(sys.argv[3], filename), 'JPEG')
shuffle(_ind)
p1size = int(len(indices) * 0.9)
indices_1, indices_2 = [indices[i] for i in sorted(_ind[:p1size])
], [indices[i] for i in sorted(_ind[p1size:])]
labels_1, labels_2 = [labels[i] for i in sorted(_ind[:p1size])
], [labels[i] for i in sorted(_ind[p1size:])]
batch_size = 128
print("Num siamese pairs", len(indices))
from keras_facenet import FaceNet
face_embedder = FaceNet()
train_embeddings = face_embedder.embeddings(train_images)
train_embeddings = normalize(train_embeddings)
generator = batch_generator(indices=indices_1,
labels=labels_1,
embeddings=train_embeddings,
batch_size=batch_size)
val = val_generator(indices=indices_2,
labels=labels_2,
embeddings=train_embeddings,
batch_size=batch_size)
def lr_schedule(epoch):
if epoch <= 30:
NAME = data[:, 515]
distance = np.array([Norm(vt, i) for i in X])
min = np.argmin(distance)
if distance[min] < 0.80:
return ID[min], NAME[min], distance[min]
return None, None, None
def is_have_face(id, name):
pass
if __name__ == "__main__":
import cv2
img = cv2.imread("trung.png")
from keras_facenet import FaceNet
embedder = FaceNet()
vt = embedder.embeddings([img])
id, name, distance = identification(vt[0])
print(id, name, distance)
Name.append(file.split(".jpg")[0])
path_file = os.path.join("person", file)
img = cv2.imread(path_file)
faces = detector.detect_faces(img)
face = faces[0]
st = (face['box'][0], face['box'][1])
en = (face['box'][0] + face['box'][2], face['box'][1] + face['box'][3])
data.append(img[st[1] : en[1], st[0]: en[0]])
img_show = img[st[1] : en[1], st[0]: en[0]]
img_show = cv2.resize(img_show, (200,200))
cv2.imshow('frame', img_show)
cv2.waitKey(1000)
vt = embedder.embeddings(data)
print(vt.shape)
df=pd.DataFrame(data=vt, index=[i for i in range(vt.shape[0] )], columns=[str(i) for i in range( vt.shape[1] )])
df['ID'] = ID
df['NAME'] = Name
df.to_csv("root.csv")
sorted_path.sort()
testing_paths = [f"{testing_path}/{path}" for path in sorted_path if path.split(".")[1] == "png"]
# cropped_testing, cropped_paths_testing = filter_faces(testing_paths)
# for img in cropped_testing:
# img = expand_dims(img, axis=3)
# print("Getting images...")
images = get_images_from_filepaths(testing_paths)
print("Getting embeddings...")
# FaceNet From keras
embedder = FaceNet()
try:
embeddings = embedder.embeddings(images, verbose=1)
except:
embeddings = embedder.embeddings(images)
dictionary_vectors = {path: embedding for path, embedding in zip(testing_paths, embeddings)}
# Split Data
names = []
for name_index in range(len(testing_paths) // 2):
names.append(basename(testing_paths[name_index * 2])[:-5])
pbar = tqdm.tqdm(total=len(names)**2)
data = {"TP": 0, "FN": 0, "TN": 0, "FP": 0}
# Distance
for name1 in names:
for name2 in names:
