def im_test(im):
face_info = lib.align(im[:, :, (2, 1, 0)], front_face_detector,
lmark_predictor)
# Samples
if len(face_info) == 0:
logging.warning('No faces are detected.')
prob = -1 # we ignore this case
else:
# Check how many faces in an image
logging.info('{} faces are detected.'.format(len(face_info)))
max_prob = -1
# If one face is fake, the image is fake
for _, point in face_info:
rois = []
for i in range(sample_num):
roi, _ = lib.cut_head([im], point, i)
rois.append(cv2.resize(roi[0], tuple(cfg.IMG_SIZE[:2])))
vis_im(rois, 'tmp/vis.jpg')
prob = solver.test(rois)
prob = np.mean(
np.sort(prob[:, 0])[np.round(sample_num / 2).astype(int):])
if prob >= max_prob:
max_prob = prob
prob = max_prob
return prob
def preprocess(im, willSimulateDeepfake=False):
'''
Given an image, attempts to detect a face. If none is found, returns None. Returns the image resized to as specified by IMG_SIZE.
If instructed to simulate Deepfake, simulation will be applied before resizing.
'''
# list of tuples of (transformation matrix, landmark point) of identified faces
faces = lib.align(im[:, :, (2, 1, 0)], front_face_detector,
lmark_predictor)
if len(faces) == 0:
return None
trans_matrix, point = faces[0] # take only the first face found
if willSimulateDeepfake:
im = simulateDeepfake(im, trans_matrix, point)
# GENERALIZATION STEP. REMOVED AFTER ACCURACY LOWERED.
# crop image, after randomly expanding ROI (minimum bounding box b) in each direction between
# [0, h/5] and [0, w/8] where h, w are height and width of b. then resize to 256 x 256 for final training data
# rois, _ = lib.cut_head([im], point, random.randint(0, 10))
# cropped_output_im = cv2.resize(rois[0], (IMG_SIZE, IMG_SIZE))
im = cv2.resize(im, (IMG_SIZE, IMG_SIZE))
return im
def preprocess(im): # refactored out of preprocess.py to accommodate the requirements of PyQt5
'''
Given an input image, preprocess it the same way as training/testing samples and return the output image.
'''
# image size = input size to model
IMG_SIZE = 256
front_face_detector = dlib.get_frontal_face_detector()
lmark_predictor = dlib.shape_predictor('./dlib_model/shape_predictor_68_face_landmarks.dat')
# list of tuples of (transformation matrix, landmark point) of identified faces
faces = lib.align(im[:, :, (2,1,0)], front_face_detector, lmark_predictor)
if len(faces) == 0:
return None
# PART REMOVED DUE TO LOWER ACCURACY
# take only the first face found
# trans_matrix, _ = faces[0]
# face = cv2.warpAffine(im, trans_matrix * FACE_SIZE, (FACE_SIZE, FACE_SIZE))
# simply resizing without cropping in to facial region, as accuracy has been found to have increased
im = cv2.resize(im, (IMG_SIZE, IMG_SIZE))
return im
def __init__(self, face_img_dir, cache_path):
self.face_img_dir = face_img_dir
self.face_img_paths = [
os.path.join(face_img_dir, im_name)
for im_name in sorted(os.listdir(face_img_dir))
]
self.data_num = len(self.face_img_paths)
self.cache_path = cache_path
# Get landmarks
face_caches = self._load_cache()
if face_caches is None:
# Load dlib
self._set_up_dlib()
face_caches = {}
# Align faces
print("Aligning faces...")
for i, im_path in enumerate(tqdm(self.face_img_paths)):
im = cv2.imread(str(im_path))
faces = lib.align(im[:, :,
(2, 1, 0)], self.front_face_detector,
self.lmark_predictor)
if len(faces) == 0:
faces = [None, None]
else:
faces = faces[0]
face_caches[self.face_img_paths[i].stem] = faces
self._save_cache(face_caches)
self.face_caches = face_caches
def im_test(net, im, args):
face_info = lib.align(im[:, :, (2, 1, 0)], front_face_detector,
lmark_predictor)
# Samples
if len(face_info) != 1:
prob = -1
else:
_, point = face_info[0]
rois = []
for i in range(sample_num):
roi, _ = lib.cut_head([im], point, i)
rois.append(cv2.resize(roi[0], (args.input_size, args.input_size)))
# vis_ = np.concatenate(rois, 1)
# cv2.imwrite('vis.jpg', vis_)
bgr_mean = np.array([103.939, 116.779, 123.68])
bgr_mean = bgr_mean[np.newaxis, :, np.newaxis, np.newaxis]
bgr_mean = torch.from_numpy(bgr_mean).float().cuda()
rois = torch.from_numpy(np.array(rois)).float().cuda()
rois = rois.permute((0, 3, 1, 2))
prob = net(rois - bgr_mean)
prob = F.softmax(prob, dim=1)
prob = prob.data.cpu().numpy()
prob = 1 - np.mean(
np.sort(prob[:, 0])[np.round(sample_num / 2).astype(int):])
return prob, face_info
def __init__(
self,
input_vid_path,
output_height=300,
):
# Input video
self.input_vid_path = input_vid_path
# parse video
print('Parsing video {}'.format(str(self.input_vid_path)))
self.imgs, self.frame_num, self.fps, self.img_w, self.img_h = pv.parse_vid(
str(self.input_vid_path))
if len(self.imgs) != self.frame_num:
warnings.warn(
'Frame number is not consistent with the number of images in video...'
)
self.frame_num = len(self.imgs)
print('Eye blinking solution is building...')
self._set_up_dlib()
self.output_height = output_height
factor = float(self.output_height) / self.img_h
# Resize imgs for final video generation
# Resize self.imgs according to self.output_height
self.aligned_imgs = []
self.left_eyes = []
self.right_eyes = []
self.resized_imgs = []
print('face aligning...')
for i, im in enumerate(tqdm(self.imgs)):
face_cache = lib.align(im[:, :,
(2, 1, 0)], self.front_face_detector,
self.lmark_predictor)
if len(face_cache) == 0:
self.left_eyes.append(None)
self.right_eyes.append(None)
continue
if len(face_cache) > 1:
raise ValueError(
'{} faces are in image, we only support one face in image.'
)
aligned_img, aligned_shapes_cur = lib.get_aligned_face_and_landmarks(
im, face_cache)
# crop eyes
leye, reye = lib.crop_eye(aligned_img[0], aligned_shapes_cur[0])
self.left_eyes.append(leye)
self.right_eyes.append(reye)
im_resized = cv2.resize(im, None, None, fx=factor, fy=factor)
self.resized_imgs.append(im_resized)
# For visualize
self.plot_vis_list = []
self.total_eye1_prob = []
self.total_eye2_prob = []