def face_swap(orig_image, down_scale):
# extract face from original
facelist = extract_faces(orig_image, 256)
result_image = orig_image
# iterate through all detected faces
for (face, resized_image) in facelist:
range_ = numpy.linspace(128 - 80, 128 + 80, 5)
mapx = numpy.broadcast_to(range_, (5, 5))
mapy = mapx.T
# warp image like in the training
mapx = mapx + numpy.random.normal(size=(5, 5), scale=5)
mapy = mapy + numpy.random.normal(size=(5, 5), scale=5)
src_points = numpy.stack([mapx.ravel(), mapy.ravel()], axis=-1)
dst_points = numpy.mgrid[0:65:16, 0:65:16].T.reshape(-1, 2)
mat = umeyama(src_points, dst_points, True)[0:2]
warped_resized_image = cv2.warpAffine(resized_image, mat,
(64, 64)) / 255.0
test_images = numpy.empty((1, ) + warped_resized_image.shape)
test_images[0] = warped_resized_image
# predict faceswap using encoder A
figure = autoencoder_A.predict(test_images)
new_face = numpy.clip(numpy.squeeze(figure[0]) * 255.0, 0,
255).astype('uint8')
mat_inv = umeyama(dst_points, src_points, True)[0:2]
# insert face into extracted face
dest_face = blend_warp(new_face, resized_image, mat_inv)
# create an inverse affine transform matrix to insert extracted face again
mat = get_align_mat(face)
mat = mat * (256 - 2 * 48)
mat[:, 2] += 48
mat_inv = cv2.invertAffineTransform(mat)
# insert new face into original image
result_image = blend_warp(dest_face, result_image, mat_inv)
# return resulting image after downscale
return cv2.resize(result_image, (result_image.shape[1] // down_scale,
result_image.shape[0] // down_scale))
def faceDetection(img, token, rid, dectype="trump"):
global graph, r
img2 = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
bounding_boxes, points = detect_face.detect_face(
img2, minsize, pnet, rnet, onet, threshold, factor)
with graph.as_default():
for box in bounding_boxes:
try:
if box[3] - box[1] < box[2] - box[0]:
delta = int((box[2] - box[0]) - (box[3] - box[1]) -
(box[2] - box[0]) * 0.2) / 2
box[3] += delta
box[1] -= delta
box[2] -= int((box[2] - box[0]) * 0.1)
box[0] += int((box[2] - box[0]) * 0.1)
else:
delta = int((box[3] - box[1]) - (box[2] - box[0]) -
(box[3] - box[1]) * 0.2) / 2
box[2] += delta
box[0] -= delta
box[3] -= int((box[3] - box[1]) * 0.1)
box[1] += int((box[3] - box[1]) * 0.1)
image = img[int(box[1]):int(box[3] + 1),
int(box[0]):int(box[2] + 1), :]
ss = image.shape
if (ss[0] * ss[1]) == 0:
continue
IMG_COL = 256
IMG_ROW = 256
border_v = 0
border_h = 0
if (IMG_COL / IMG_ROW) >= (image.shape[0] /
image.shape[1]):
border_v = int(
(((IMG_COL / IMG_ROW) * image.shape[1]) -
image.shape[0]) / 2)
image = cv2.copyMakeBorder(image, border_v, border_v,
0, 0, cv2.BORDER_REPLICATE,
0)
else:
border_h = int(
(((IMG_ROW / IMG_COL) * image.shape[0]) -
image.shape[1]) / 2)
image = cv2.copyMakeBorder(image, 0, 0, border_h,
border_h,
cv2.BORDER_REPLICATE, 0)
except Exception as e:
print(e)
continue
try:
ss = image.shape
#print(ss)
image = cv2.resize(image, (IMG_ROW, IMG_COL)) / 255.0
#cv2.normalize(image, image, 0, 1.0, norm_type=cv2.NORM_MINMAX)
test = np.empty((1, ) + image.shape, image.dtype)
test[0] = image
if (dectype == "trump"):
figure = autoencoder_A.predict(test)
elif (dectype == "swift"):
figure = autoencoder_A_swift.predict(test)
else:
figure = autoencoder_B.predict(test)
a1 = cv2.resize(alpha, (ss[1], ss[0]))
figure = cv2.resize(
figure[0, :, :, :],
(ss[1], ss[0])) #[border_h:-border_h,:,:]
figure = cv2.filter2D(
figure, -1,
np.array([[-1, -1, -1], [-1, 9, -1], [-1, -1, -1]]))
figure = np.clip(figure * 255.0, 0, 255).astype('uint8')
img[int(box[1]):int(box[1]) + ss[0],
int(box[0]
):int(box[0] + ss[1]), :] = cv2.convertScaleAbs(
img[int(box[1]):int(box[1]) + ss[0],
int(box[0]):int(box[0] + ss[1]), :] *
(1 - a1) + figure * a1)
except ValueError as e:
print(e)
jpg = cv2.imencode(".jpg", img)[1].tostring()
rid = rid + ":face"
r.set(rid, jpg, ex=30)
r.publish(token + "face", rid)
#print(token,rid)
return
test_B_i = numpy.array(test_B_i).reshape((-1,64,64,3))
figWarped = numpy.stack([warped_A[:6],warped_B[:6]],axis=0 )
figWarped = numpy.clip( figWarped * 255, 0, 255 ).astype('uint8')
figWarped = stack_images( figWarped )
cv2.imshow( "w", figWarped )
zmask = numpy.zeros((test_A.shape[0],128,128,1),float)
pred_a_a,pred_a_a_m = autoencoder_A.predict([test_A_i,zmask])
pred_b_a,pred_b_a_m = autoencoder_B.predict([test_A_i,zmask])
pred_a_b,pred_a_b_m = autoencoder_A.predict([test_B_i,zmask])
pred_b_b,pred_b_b_m = autoencoder_B.predict([test_B_i,zmask])
pred_a_a = pred_a_a[0:18,:,:,:3]
pred_a_b = pred_a_b[0:18,:,:,:3]
pred_b_a = pred_b_a[0:18,:,:,:3]
pred_b_b = pred_b_b[0:18,:,:,:3]
figure_A = numpy.stack([
test_A,
pred_a_a,
pred_b_a,
], axis=1 )
batch_size = 64
warped_A, target_A = get_training_data( images_A, batch_size )
warped_B, target_B = get_training_data( images_B, batch_size )
loss_A = autoencoder_A.train_on_batch( warped_A, target_A )
loss_B = autoencoder_B.train_on_batch( warped_B, target_B )
print( loss_A, loss_B )
if epoch % 100 == 0:
save_model_weights()
test_A = target_A[0:14]
test_B = target_B[0:14]
figure_A = numpy.stack([
test_A,
autoencoder_A.predict( test_A ),
autoencoder_B.predict( test_A ),
], axis=1 )
figure_B = numpy.stack([
test_B,
autoencoder_B.predict( test_B ),
autoencoder_A.predict( test_B ),
], axis=1 )
figure = numpy.concatenate( [ figure_A, figure_B ], axis=0 )
figure = figure.reshape( (4,7) + figure.shape[1:] )
figure = stack_images( figure )
figure = numpy.clip( figure * 255, 0, 255 ).astype('uint8')
cv2.imshow( "", figure )
wrap ,a_faces = get_training_data(imgsA,6)
wrap2 ,b_faces = get_training_data(imgsB,6)
#input()
# show original image
for (index, img) in enumerate(a_faces):
winn="original_image1_"+str(index)
cv2.namedWindow(winn)
cv2.moveWindow(winn, 10,index*170)
cv2.imshow(winn, img)
#a_faces = a_faces.astype('float32') / 255.
#wrap = wrap.astype('float32') / 255.
decoded_imgs = autoencoder_A.predict(a_faces)
#decoded_imgs = (decoded_imgs *255).astype(np.uint8)
print(decoded_imgs.shape)
for (index, img) in enumerate(decoded_imgs):
winn="dec_image1_"+str(index)
cv2.namedWindow(winn)
cv2.moveWindow(winn, 130,index*170)
cv2.imshow(winn, img)
decoded_imgs = autoencoder_B.predict(a_faces)
#decoded_imgs = (decoded_imgs*255).astype(np.uint8)
for (index, img) in enumerate(decoded_imgs):
winn="dec_image4_"+str(index)
cv2.namedWindow(winn)
cv2.moveWindow(winn, 250,index*170)
cv2.imshow(winn, img)
else:
offset = int((face_h - face_w) / 2)
face_patch = face_patch[offset:offset + face_w, :, :]
# Crop face patch
offset_crop = face_patch.shape[0] // 10
face_patch_crop = face_patch[offset_crop:face_patch.shape[0] -
offset_crop,
offset_crop:face_patch.shape[1] -
offset_crop, :]
frame_test = cv2.resize(face_patch_crop, (64, 64),
interpolation=cv2.INTER_AREA)
frame_test = frame_test / 255.0
# Inference
frame_hat = autoencoder_A.predict(frame_test.reshape((1, 64, 64, 3)))
frame_hat = frame_hat[0, :, :, :]
frame_hat = np.clip(frame_hat * 255, 0, 255).astype('uint8')
# Stitch
face_stitch = cv2.resize(
frame_hat, (face_patch_crop.shape[0], face_patch_crop.shape[1]),
interpolation=cv2.INTER_AREA)
frame_stitch = np.copy(frame)
if face_h < face_w:
x_start = face_x + offset_crop
x_end = face_x + face_w - offset_crop
y_start = face_y + offset + offset_crop
y_end = face_y + offset + face_h - offset_crop
else:
batch_size = 64
warped_A, target_A = get_training_data( images_A, batch_size )
warped_B, target_B = get_training_data( images_B, batch_size )
loss_A = autoencoder_A.train_on_batch( warped_A, target_A )
loss_B = autoencoder_B.train_on_batch( warped_B, target_B )
print( loss_A, loss_B )
if epoch % 100 == 0:
save_model_weights()
test_A = target_A[0:14]
test_B = target_B[0:14]
figure_A = numpy.stack([
test_A,
autoencoder_A.predict( test_A ),
autoencoder_B.predict( test_A ),
], axis=1 )
figure_B = numpy.stack([
test_B,
autoencoder_B.predict( test_B ),
autoencoder_A.predict( test_B ),
], axis=1 )
figure = numpy.concatenate( [ figure_A, figure_B ], axis=0 )
figure = figure.reshape( (4,7) + figure.shape[1:] )
figure = stack_images( figure )
figure = numpy.clip( figure * 255, 0, 255 ).astype('uint8')
cv2.imshow( "", figure )