Face Recognition From Image Blog

compare-face-encodings.py

# TOLERANCE defines the tolerance for face comparisons
# The lower the number - the stricter the comparison
# To avoid false matches, use lower value
# To avoid false negatives (i.e. faces of the same person doesn't match), use higher value
# 0.5-0.6 works well
TOLERANCE = 0.50

# This function returns the name of the person whose image matches with the given face (or 'Unknown Face')
# face_encoding is the face encoding of the face we are looking for
def find_match(face_encoding):

    # see if the face is any match(es) for the known face(s)
    # Compute the face distance to get a list of True/False values indicating whether or not there's a match
    all_matches = (np.linalg.norm(known_face_encodings - face_encoding, axis=1) <= TOLERANCE).tolist()
    
    # initialize name string as Unknown Face
    name = 'Unknown Face'

    # if a match was found in known_face_encodings, use the first one.
    if True in all_matches:
        first_match_index = all_matches.index(True)
        name = known_face_names[first_match_index]

    # Return the name of the first match
    return name

dir-tree.md

•
├── images
│   ├── training
│   │   ├── Barack Obama.jpg
│   │   ├── Narendra Modi.jpg
│   │   ├── Boris Johnson.jpg
│   │   ├── Donald Trump.jpg
│   │   └── Justin Trudeau.jpg
│   └── testing
│       ├── modi-obama-1.jpg
│       ├── modi-obama-2.jpg
│       ├── trump-modi.jpg
│       └── trump-obama.png
├── models
│   ├── mmod_human_face_detector.dat
│   ├── shape_predictor_68_face_landmarks.dat
│   └── dlib_face_recognition_resnet_model_v1.dat
├── dlib-face-recognition-image.py
└── face-recognition-from-image.py

dlib-box-to-bb.py

# convert dlib's bounding box object to (top_pos, right_pos, bottom_pos, left_pos) tuple
def dlib_box_to_bb(face_location):
    
    left_pos = face_location.left()
    top_pos = face_location.top()
    right_pos = face_location.right()
    bottom_pos = face_location.bottom()

    # return face location tupple
    return top_pos, right_pos, bottom_pos, left_pos

dlib-face-recognition.py

# read and load an image file
image_to_recognize = cv2.imread('images/testing/modi-obama-2.jpg')

# apply face detector and localize all human faces from the image
# if HOG face detector is used
# all_face_locations = hog_face_detector(image_to_recognize, 1)

# if CNN face detector is used
all_face_locations = cnn_face_detector(image_to_recognize, 1)

for index, face_location in enumerate(all_face_locations):

    # if HOG face detector is used
    # face_location = face_location

    # if CNN face detector is used
    face_location = face_location.rect

    # align the face and calculate face pos using dlib's shape predictor
    face_pose = shape_predictor(image_to_recognize, face_location)    

    # compute encodings and save as numpy array
    face_encoding = np.array(dlib_face_recognition_model.compute_face_descriptor(image_to_recognize, face_pose, 1))

    name = find_match(face_encoding)
    
    # convert dlib's bounding box object to (top_pos, right_pos, bottom_pos, left_pos)
    # if HOG face detector is used
    # top_pos, right_pos, bottom_pos, left_pos = dlib_box_to_bb(face_location)

    # if CNN face detector is used
    top_pos, right_pos, bottom_pos, left_pos = dlib_box_to_bb(face_location)

    # draw rectangle around face
    cv2.rectangle(image_to_recognize, (left_pos,top_pos), (right_pos,bottom_pos), (0,0,255), 2)

    # disply the name of age as text along the face
    font = cv2.FONT_HERSHEY_SIMPLEX
    cv2.putText(image_to_recognize, name, (left_pos,bottom_pos+15), font, 0.5, (0,0,255), 1)

# show image
cv2.imshow("Image Face Recognition",image_to_recognize)

# only for google colab 
# cv2_imshow(image_to_recognize)

face-recognition-from-image.py

# importing required libraries
import cv2
import  face_recognition
import os

# only for google colab
# from google.colab.patches import cv2_imshow

# initialize the empty list of name label
known_face_names = []

# initialize the empty list for storing the encoding of each face
known_face_encodings = []

# initialize the folder path consisting of training dataset
training_dataset_path = 'images/training/'

# read all the image file with extension .JPG present under the training_dataset_path
training_dataset = filter(lambda x: x.endswith('.jpg'), os.listdir(training_dataset_path))

# Sort in alphabetical order
training_dataset = sorted(training_dataset)

# define function for generating known face encodings
def generate_known_face_encodings():

  # Get full paths to training images
  paths_to_training_images = [training_dataset_path + x for x in training_dataset]

  for index,image_file in enumerate(paths_to_training_images):

    # Get list of names of people by eliminating the .JPG extension from image filenames
    known_face_names.append(image_file[image_file.rindex('/')+1:][:-4])

    # read image file
    image = cv2.imread(image_file)
    
    # detect face from image and get the face encoding
    face_encoding = face_recognition.face_encodings(image)[0]

    # append this into known_face_encodings
    known_face_encodings.append(face_encoding)
    
print("Training Started...")
generate_known_face_encodings()
print("Training Completed...")

# load the image to be detected
image_to_recognize = cv2.imread('modi-obama-1.jpg')

# locate all faces from the image
all_face_locations = face_recognition.face_locations(image_to_recognize, number_of_times_to_upsample = 2 ,model="cnn")

# generate encoding for all detected faces
all_face_encodings = face_recognition.face_encodings(image_to_recognize,all_face_locations)

for face_location, face_encoding in zip(all_face_locations,all_face_encodings):

  # split the tuple
  top_pos, right_pos, bottom_pos, left_pos = face_location

  # see if the face is any match(es) for the known face(s)
  all_matches = face_recognition.compare_faces(known_face_encodings,face_encoding)

  # initialize a name string as unknown face
  name = "Unknown Face"

  # if a match was found in known_face_encodings, use the first one.
  if True in all_matches:
    first_match_index = all_matches.index(True)
    name = known_face_names[first_match_index]

  # draw rectangle around face
  cv2.rectangle(image_to_recognize, (left_pos,top_pos), (right_pos,bottom_pos), (0,0,255), 2)

  # disply the name of age as text along the face
  font = cv2.FONT_HERSHEY_SIMPLEX
  cv2.putText(image_to_recognize, name, (left_pos,bottom_pos+15), font, 0.5, (0,0,255), 1)

# show image
cv2.imshow("Image Face Recognition",image_to_recognize)

# only for google colab 
# cv2_imshow(image_to_recognize)

import-libraries.py

# inport required libraries
import cv2
import dlib
import os
import numpy as np

# only for google colab
# from google.colab.patches import cv2_imshow

initialize-models-and-weights.py

# initialize face detector model
# dlib's HOG detector
hog_face_detector = dlib.get_frontal_face_detector()

# dlib's CNN detector
detector_weights = 'models/mmod_human_face_detector.dat'
cnn_face_detector = dlib.cnn_face_detection_model_v1(detector_weights)

# initialize dlib shape predictor
predictor_path = 'models/shape_predictor_68_face_landmarks.dat'
shape_predictor = dlib.shape_predictor(predictor_path)

# initialize dlib face recognition
recognition_weigths = 'models/dlib_face_recognition_resnet_model_v1.dat'
dlib_face_recognition_model = dlib.face_recognition_model_v1(recognition_weigths)

train-model-on-labeled-images.py

# initialize the empty list of name label
known_face_names = []

# initialize the empty list for storing the encoding of each face
known_face_encodings = []

# initialize the folder path consisting of training dataset
training_dataset_path = 'images/training/'

# read all the image file with extension .JPG present under the training_dataset_path
training_dataset = filter(lambda x: x.endswith('.jpg'), os.listdir(training_dataset_path))

# Sort in alphabetical order
training_dataset = sorted(training_dataset)

# define function for generating known face encodings
def generate_known_face_encodings():

  # Get full paths to training images
  paths_to_training_images = [training_dataset_path + x for x in training_dataset]

  for index,image_file in enumerate(paths_to_training_images):

    # read image file
    image = cv2.imread(image_file)

    # apply face detector and localize human faces from the image
    # if HOG face detector is used
    # all_face_locations = hog_face_detector(image, 1)
    
    # if CNN face detector is used
    all_face_locations = cnn_face_detector(image, 1)

    # verify that each training image contains only 1 face
    if len(all_face_locations) != 1:
        print(image_file,', must contain only 1 face while training.')
        continue

    #--------------------------------------------------------------------------
    # generate encodings
    #--------------------------------------------------------------------------

    # convert all_face_locations[0] to dlib rectangle object

    # if HOG face detector is used
    # face_location = all_face_locations[0]

    # if CNN face detector is used
    face_location = all_face_locations[0].rect

    # align the face and calculate face pos using dlib's shape predictor
    face_pose = shape_predictor(image, face_location)    

    # compute encodings and save as numpy array
    face_encoding = np.array(dlib_face_recognition_model.compute_face_descriptor(image, face_pose, 1))

    # append to known_face_encodings[]
    known_face_encodings.append(face_encoding)

    # Get list of names of people by eliminating the .JPG extension from image filenames
    known_face_names.append(image_file[image_file.rindex('/')+1:][:-4])

    #--------------------------------------------------------------------------

print("Training Started...")
generate_known_face_encodings()
print("Training Completed...")