# Path to landmarks and face recognition model files
PREDICTOR_PATH = '../resource/lib/publicdata/models/shape_predictor_68_face_landmarks.dat'
FACE_RECOGNITION_MODEL_PATH = '../resource/lib/publicdata/models/dlib_face_recognition_resnet_model_v1.dat'
# Initialize face detector, facial landmarks detector
# and face recognizer
faceDetector = dlib.get_frontal_face_detector()
shapePredictor = dlib.shape_predictor(PREDICTOR_PATH)
faceRecognizer = dlib.face_recognition_model_v1(FACE_RECOGNITION_MODEL_PATH)
# Root folder of the dataset
faceDatasetFolder = '../resource/asnlib/publicdata/celeb_mini'
# Label -> Name Mapping file
labelMap = np.load("../resource/asnlib/publicdata/celeb_mapping.npy", allow_pickle=True).item()
{'n00000001': 'A.J. Buckley', 'n00000002': 'A.R. Rahman',...}
We find every subfolder has 5 images. So we will create dictionary named "nameLabelMap' for each images
imagePaths = []
nameLabelMap = {}
Labels = []
celenames=[]
subfolders = []
for x in os.listdir(faceDatasetFolder):
xpath = os.path.join(faceDatasetFolder, x)
if os.path.isdir(xpath):
subfolders.append(xpath)
for i, subfolder in enumerate(subfolders):
for x in os.listdir(subfolder):
xpath = os.path.join(subfolder, x)
if x.endswith('JPEG'):
imagePaths.append(xpath)
Labels.append(i)
celenames.append(labelMap[xpath.split('/')[-2]])
nameLabelMap=dict(zip(imagePaths,celenames))
4.1 detect faces in image:
index = {}
pathidx=[] # save detected face image path
i = 0
faceDescriptors = None
for imagePath in imagePaths:
print("processing: {}".format(imagePath))
img = cv2.imread(imagePath)
faces = faceDetector(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))#BGR to RGB
print("{} Face(s) found".format(len(faces)))
4.2 Now process each face we found:
for k, face in enumerate(faces):
#Find facial landmarks for each detected face
shape = shapePredictor(cv2.cvtColor(img, cv2.COLOR_BGR2RGB), face)
# convert landmarks from Dlib's format to list of (x, y) points
landmarks = [(p.x, p.y) for p in shape.parts()]
# Compute face descriptor using neural network defined in Dlib.
# It is a 128D vector that describes the face in img identified by shape.
faceDescriptor = faceRecognizer.compute_face_descriptor(img, shape)
# Convert face descriptor from Dlib's format to list, then a NumPy array
faceDescriptorList = [x for x in faceDescriptor]
faceDescriptorNdarray = np.asarray(faceDescriptorList, dtype=np.float64)
faceDescriptorNdarray = faceDescriptorNdarray[np.newaxis, :]
# Stack face descriptors (1x128) for each face in images, as rows
if faceDescriptors is None:
faceDescriptors = faceDescriptorNdarray
else:
faceDescriptors = np.concatenate((faceDescriptors, faceDescriptorNdarray), axis=0)
# save the label for this face in index. We will use it later to identify
# person name corresponding to face descriptors stored in NumPy Array
index[i] = nameLabelMap[imagePath]
i += 1
# save the fullpath for each face-detected image in same order. We will show the looklike person later.
pathidx.append(imagePath)
4.3 saving detection files in disk, we will load and use in test module code
np.save('descriptors.npy', faceDescriptors)
with open('index.pkl', 'wb') as f:
cPickle.dump(index, f)
5.2 compare the distances between test image and every enrolled image in descriptors.npy saved last step.
5.3 get the minimun distance's index in distances list, we don't need the value of threshold, just get the most likely person's index.
# read test image
testImages = glob.glob('../resource/asnlib/publicdata/test-images/*.jpg')
for test in testImages:
im = cv2.imread(test)
imDlib = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
celeb_name = ""
faceDetector = dlib.get_frontal_face_detector()
shapePredictor = dlib.shape_predictor(PREDICTOR_PATH)
faceRecognizer = dlib.face_recognition_model_v1(FACE_RECOGNITION_MODEL_PATH)
#load the name of enrolled image in order
index = np.load('index.pkl', allow_pickle=True)
#load enrolled image in order
faceDescriptorsEnrolled = np.load('descriptors.npy')
#face detection in test image
img = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
faces = faceDetector(cv2.cvtColor(im, cv2.COLOR_BGR2RGB))
#compare test image with all enrolled image
for face in faces:
shape = shapePredictor(cv2.cvtColor(im, cv2.COLOR_BGR2RGB), face)
faceDescriptor = faceRecognizer.compute_face_descriptor(im, shape)
faceDescriptorList = [m for m in faceDescriptor]
faceDescriptorNdarray = np.asarray(faceDescriptorList, dtype=np.float64)
faceDescriptorNdarray = faceDescriptorNdarray[np.newaxis, :]
#caculate distances between test image and enrolled image
distances = np.linalg.norm(faceDescriptorsEnrolled - faceDescriptorNdarray, axis=1)
#find the most likely person's index
argmin = np.argmin(distances)
# minimum distance
minDistance = distances[argmin]
# find the most likely person's name and full path
label = index[argmin]
#the name of most look like person.
celeb_name=index[argmin]
#the full path of most look like person image
pathout=pathidx[argmin]
plt.subplot(121)
plt.imshow(imDlib)
plt.title("test img")
#TODO - display celeb image which looks like the test image instead of the black image.
plt.subplot(122)
likeimg=cv2.imread(pathout)
likeimgDlib=cv2.cvtColor(likeimg, cv2.COLOR_BGR2RGB)
plt.imshow(likeimgDlib)
plt.title("Celeb Look-Alike={}".format(celeb_name))
plt.show()
##the results: