def train_model_CNN_LSTM_New(files_original, files_fake):
files = []
files.extend(files_original)
files.extend(files_fake)
np.random.shuffle(files)
CNN_LSTM_model = getCustomCNNLSTMModel(number_of_faces, height, width, 3)
for file in files:
frames = get_frames(file, startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
faces = np.asarray(faces)
faces = np.array(faces)
if (len(faces) != number_of_faces):
continue
if (file in files_original):
label = [0, 1]
if (file in files_fake):
label = [1, 0]
input_for_LSTM = faces.reshape(1, number_of_faces, height, width, 3)
labels = []
labels.append(label)
y_train = np.asarray(labels)
CNN_LSTM_model.fit(input_for_LSTM, y_train, epochs=20, verbose=1)
del faces
del input_for_LSTM
del frames
CNN_LSTM_model.save('CNN_New_lstmModel.h5')
def test_model_CNN_Lstm(model_name, files):
model = load_model(model_name)
for file in files:
print("File")
frames = get_frames(file, number_of_frames=-1, startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
# if(len(faces)!=number_of_faces):
# print("File: " + str(file))
# print("No Result Found\n")
# continue;
# else:
# tempFaces.extend(faces)
# count = count+1;
testFaces = np.asarray(faces)
input_for_LSTM = testFaces.reshape(1, number_of_faces, height, width,
3)
print("Shape for testing:", input_for_LSTM.shape)
y_test_result = model.predict(input_for_LSTM)
print("\n\n")
print("Result: ")
print(y_test_result)
print("\n")
def get_faces_local_for_CNN(files_original, files_fake):
np.set_printoptions(threshold=sys.maxsize)
tempFaces = []
labels = []
for original_file in files_original:
frames = get_frames(original_file, startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
if (len(faces) == number_of_faces):
tempFaces.extend(faces)
for face in faces:
labels.append([0, 1])
del frames
del faces
facesCorrect = np.asarray(tempFaces)
tempFaces = []
for fake_file in files_fake:
frames = get_frames(fake_file, startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
if (len(faces) == number_of_faces):
tempFaces.extend(faces)
for face in faces:
labels.append([1, 0])
del frames
del faces
facesIncorrect = np.asarray(tempFaces)
count_incorrect = len(facesIncorrect)
count_correct = len(facesCorrect)
print(facesIncorrect.shape)
x_train = np.concatenate((facesCorrect, facesIncorrect))
return x_train, count_incorrect, count_correct, labels
def test_model(files):
model = load_model('classification_CNN.h5')
for file in files:
frames = get_frames(file, number_of_frames=40, startingPoint=0)
tempFaces = (get_faces(frames, height=height, width=width))
testFaces = np.asarray(tempFaces)
testFaces_shape = testFaces.shape
testFaces = testFaces.reshape(testFaces_shape[0], testFaces_shape[1],
testFaces_shape[2], 1)
print(testFaces[0].shape)
#testFaces = np.concatenate(testFaces)
print(testFaces.shape)
y_test_result = model.predict(testFaces)
count_fake = 0
count_positive = 0
count_clear_fake = 0
count_clear_positive = 0
for frame_result in y_test_result:
if (frame_result[0] - frame_result[1] >= 0.3):
count_clear_positive = count_clear_positive + 1
elif (frame_result[0] - frame_result[1] > 0):
count_positive = count_positive + 1
elif (frame_result[1] - frame_result[0] >= 0.3):
count_clear_fake = count_clear_fake + 1
elif (frame_result[1] - frame_result[0] > 0):
count_fake = count_fake + 1
else:
count_doubt = count_doubt + 1
print("File: " + file)
print("count_clear positive:" + str(count_clear_positive))
print("count_positive: " + str(count_positive))
print("count_fake: " + str(count_fake))
print("count_clear_fake: " + str(count_clear_fake))
print("\n")
del frames
height = 320
width=320
number_of_faces=40
def get_all_files(folder):
filepaths = [os.path.join(folder, f) for f in os.listdir(folder)]
return filepaths
files_fake = get_all_files('../../manipulated_sequences/Deepfakes/raw/videos/')
files_original = get_all_files('../../original_sequences/youtube/raw/videos/')
for file in files_fake:
frames = get_frames(file, startingPoint=0)
faces = get_faces(frames, height=height, width=width, number_of_faces=1)
if(len(faces)>0):
photo = cv.Canny(faces[0],0,10)
fig, ax = plt.subplots(1, 1, figsize=(15, 15))
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
plt.grid(False)
ax.imshow(photo)
plt.show()
# fig, ax = plt.subplots(nrows=2, ncols=3, figsize=(8, 5),
# sharex=True, sharey=True)
from face_detect_util.get_face import get_frames, get_faces, get_cropped_images
height = 320
width = 320
number_of_faces = 2
def get_all_files(folder):
filepaths = [os.path.join(folder, f) for f in os.listdir(folder)]
return filepaths
files_fake = get_all_files('../../manipulated_sequences/Deepfakes/raw/videos/')
files_original = get_all_files('../../original_sequences/youtube/raw/videos/')
frames = get_frames(files_fake[2], startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
original_true = img_as_float(faces[0])
frames = get_frames(files_original[0], startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
original_false = img_as_float(faces[0])
sigma = 0.12
noisy_true = random_noise(original_true, var=sigma**2)
height = 299
width = 299
number_of_faces = 40
def get_all_files(folder):
filepaths = [os.path.join(folder, f) for f in os.listdir(folder)]
return filepaths
files_fake = get_all_files('../manipulated_sequences/Deepfakes/raw/videos/')
files_original = get_all_files('../original_sequences/youtube/raw/videos/')
for file_fake, file_original in zip(files_fake, files_original):
frames = get_frames(file_fake, startingPoint=0)
faces = get_faces(frames,
height=height,
width=width,
number_of_faces=number_of_faces)
if (len(faces) == number_of_faces):
for face in faces:
fig, ax = plt.subplots(1, 1, figsize=(15, 15))
ax.imshow(face)
ax.xaxis.set_visible(False)
ax.yaxis.set_visible(False)
ax.title.set_text(f"FRAME 0: {file_fake.split('/')[-1]}")
plt.grid(False)
plt.show()
frames = get_frames(file_original, startingPoint=0)