github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-07-12 08:07:03 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

Added interactive converter.

Browse source 
With interactive converter you can change any parameter of any frame and see the result in real time.

Converter: added motion_blur_power param.
Motion blur is applied by precomputed motion vectors.
So the moving face will look more realistic.

RecycleGAN model is removed.

Added experimental AVATAR model. Minimum required VRAM is 6GB (NVIDIA), 12GB (AMD)
Usage:
1) place data_src.mp4 10-20min square resolution video of news reporter sitting at the table with static background,
   other faces should not appear in frames.
2) process "extract images from video data_src.bat" with FULL fps
3) place data_dst.mp4 video of face who will control the src face
4) process "extract images from video data_dst FULL FPS.bat"
5) process "data_src mark faces S3FD best GPU.bat"
6) process "data_dst extract unaligned faces S3FD best GPU.bat"
7) train AVATAR.bat stage 1, tune batch size to maximum for your card (32 for 6GB), train to 50k+ iters.
8) train AVATAR.bat stage 2, tune batch size to maximum for your card (4 for 6GB), train to decent sharpness.
9) convert AVATAR.bat
10) converted to mp4.bat

updated versions of modules
This commit is contained in:
iperov 2019-08-24 12:57:29 +04:00
parent 3f0bf2e994
commit 407ce3b1ca
46 changed files with 2394 additions and 1659 deletions
Download patch file Download diff file Expand all files Collapse all files

392
converters/ConvertMasked.py Normal file
View file

@ -0,0 +1,392 @@
import traceback
import cv2
import numpy as np
import imagelib
from facelib import FaceType, LandmarksProcessor
from interact import interact as io
from utils.cv2_utils import *
def ConvertMaskedFace (cfg, frame_info, img_bgr_uint8, img_bgr, img_face_landmarks):
#if debug:
# debugs = [img_bgr.copy()]
img_size = img_bgr.shape[1], img_bgr.shape[0]
img_face_mask_a = LandmarksProcessor.get_image_hull_mask (img_bgr.shape, img_face_landmarks)
if cfg.mode == 'original':
if cfg.export_mask_alpha:
img_bgr = np.concatenate ( [img_bgr, img_face_mask_a], -1 )
return img_bgr, img_face_mask_a
out_img = img_bgr.copy()
out_merging_mask = None
output_size = cfg.predictor_input_shape[0]
if cfg.super_resolution:
output_size *= 2
face_mat = LandmarksProcessor.get_transform_mat (img_face_landmarks, output_size, face_type=cfg.face_type)
face_output_mat = LandmarksProcessor.get_transform_mat (img_face_landmarks, output_size, face_type=cfg.face_type, scale= 1.0 + 0.01*cfg.output_face_scale )
dst_face_bgr = cv2.warpAffine( img_bgr , face_mat, (output_size, output_size), flags=cv2.INTER_CUBIC )
dst_face_mask_a_0 = cv2.warpAffine( img_face_mask_a, face_mat, (output_size, output_size), flags=cv2.INTER_CUBIC )
predictor_input_bgr = cv2.resize (dst_face_bgr, cfg.predictor_input_shape[0:2] )
if cfg.predictor_masked:
prd_face_bgr, prd_face_mask_a_0 = cfg.predictor_func (predictor_input_bgr)
prd_face_bgr = np.clip (prd_face_bgr, 0, 1.0 )
prd_face_mask_a_0 = np.clip (prd_face_mask_a_0, 0.0, 1.0)
else:
predicted = cfg.predictor_func (predictor_input_bgr)
prd_face_bgr = np.clip (predicted, 0, 1.0 )
prd_face_mask_a_0 = cv2.resize (dst_face_mask_a_0, cfg.predictor_input_shape[0:2] )
if cfg.super_resolution:
#if debug:
# tmp = cv2.resize (prd_face_bgr, (output_size,output_size), cv2.INTER_CUBIC)
# debugs += [ np.clip( cv2.warpAffine( tmp, face_output_mat, img_size, img_bgr.copy(), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
prd_face_bgr = cfg.dcscn_upscale_func(prd_face_bgr)
#if debug:
# debugs += [ np.clip( cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, img_bgr.copy(), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
if cfg.predictor_masked:
prd_face_mask_a_0 = cv2.resize (prd_face_mask_a_0, (output_size, output_size), cv2.INTER_CUBIC)
else:
prd_face_mask_a_0 = cv2.resize (dst_face_mask_a_0, (output_size, output_size), cv2.INTER_CUBIC)
if cfg.mask_mode == 2: #dst
prd_face_mask_a_0 = cv2.resize (dst_face_mask_a_0, (output_size,output_size), cv2.INTER_CUBIC)
elif cfg.mask_mode >= 3 and cfg.mask_mode <= 7:
if cfg.mask_mode == 3 or cfg.mask_mode == 5 or cfg.mask_mode == 6:
prd_face_fanseg_bgr = cv2.resize (prd_face_bgr, (cfg.fanseg_input_size,)*2 )
prd_face_fanseg_mask = cfg.fanseg_extract_func(FaceType.FULL, prd_face_fanseg_bgr)
FAN_prd_face_mask_a_0 = cv2.resize ( prd_face_fanseg_mask, (output_size, output_size), cv2.INTER_CUBIC)
if cfg.mask_mode >= 4 or cfg.mask_mode <= 7:
full_face_fanseg_mat = LandmarksProcessor.get_transform_mat (img_face_landmarks, cfg.fanseg_input_size, face_type=FaceType.FULL)
dst_face_fanseg_bgr = cv2.warpAffine(img_bgr, full_face_fanseg_mat, (cfg.fanseg_input_size,)*2, flags=cv2.INTER_CUBIC )
dst_face_fanseg_mask = cfg.fanseg_extract_func( FaceType.FULL, dst_face_fanseg_bgr )
if cfg.face_type == FaceType.FULL:
FAN_dst_face_mask_a_0 = cv2.resize (dst_face_fanseg_mask, (output_size,output_size), cv2.INTER_CUBIC)
elif cfg.face_type == FaceType.HALF:
half_face_fanseg_mat = LandmarksProcessor.get_transform_mat (img_face_landmarks, cfg.fanseg_input_size, face_type=FaceType.HALF)
fanseg_rect_corner_pts = np.array ( [ [0,0], [cfg.fanseg_input_size-1,0], [0,cfg.fanseg_input_size-1] ], dtype=np.float32 )
a = LandmarksProcessor.transform_points (fanseg_rect_corner_pts, half_face_fanseg_mat, invert=True )
b = LandmarksProcessor.transform_points (a, full_face_fanseg_mat )
m = cv2.getAffineTransform(b, fanseg_rect_corner_pts)
FAN_dst_face_mask_a_0 = cv2.warpAffine(dst_face_fanseg_mask, m, (cfg.fanseg_input_size,)*2, flags=cv2.INTER_CUBIC )
FAN_dst_face_mask_a_0 = cv2.resize (FAN_dst_face_mask_a_0, (output_size,output_size), cv2.INTER_CUBIC)
else:
raise ValueError ("cfg.face_type unsupported")
if cfg.mask_mode == 3: #FAN-prd
prd_face_mask_a_0 = FAN_prd_face_mask_a_0
elif cfg.mask_mode == 4: #FAN-dst
prd_face_mask_a_0 = FAN_dst_face_mask_a_0
elif cfg.mask_mode == 5:
prd_face_mask_a_0 = FAN_prd_face_mask_a_0 * FAN_dst_face_mask_a_0
elif cfg.mask_mode == 6:
prd_face_mask_a_0 = prd_face_mask_a_0 * FAN_prd_face_mask_a_0 * FAN_dst_face_mask_a_0
elif cfg.mask_mode == 7:
prd_face_mask_a_0 = prd_face_mask_a_0 * FAN_dst_face_mask_a_0
prd_face_mask_a_0[ prd_face_mask_a_0 < 0.001 ] = 0.0
prd_face_mask_a = prd_face_mask_a_0[...,np.newaxis]
prd_face_mask_aaa = np.repeat (prd_face_mask_a, (3,), axis=-1)
img_face_mask_aaa = cv2.warpAffine( prd_face_mask_aaa, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), flags=cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC )
img_face_mask_aaa = np.clip (img_face_mask_aaa, 0.0, 1.0)
img_face_mask_aaa [ img_face_mask_aaa <= 0.1 ] = 0.0 #get rid of noise
#if debug:
# debugs += [img_face_mask_aaa.copy()]
if 'raw' in cfg.mode:
face_corner_pts = np.array ([ [0,0], [output_size-1,0], [output_size-1,output_size-1], [0,output_size-1] ], dtype=np.float32)
square_mask = np.zeros(img_bgr.shape, dtype=np.float32)
cv2.fillConvexPoly(square_mask, \
LandmarksProcessor.transform_points (face_corner_pts, face_output_mat, invert=True ).astype(np.int), \
(1,1,1) )
if cfg.mode == 'raw-rgb':
out_merging_mask = square_mask
if cfg.mode == 'raw-rgb' or cfg.mode == 'raw-rgb-mask':
out_img = cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, out_img, cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT )
if cfg.mode == 'raw-rgb-mask':
out_img = np.concatenate ( [out_img, np.expand_dims (img_face_mask_aaa[:,:,0],-1)], -1 )
out_merging_mask = square_mask
elif cfg.mode == 'raw-mask-only':
out_img = img_face_mask_aaa
out_merging_mask = img_face_mask_aaa
elif cfg.mode == 'raw-predicted-only':
out_img = cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT )
out_merging_mask = square_mask
out_img = np.clip (out_img, 0.0, 1.0 )
else:
#averaging [lenx, leny, maskx, masky] by grayscale gradients of upscaled mask
ar = []
for i in range(1, 10):
maxregion = np.argwhere( img_face_mask_aaa > i / 10.0 )
if maxregion.size != 0:
miny,minx = maxregion.min(axis=0)[:2]
maxy,maxx = maxregion.max(axis=0)[:2]
lenx = maxx - minx
leny = maxy - miny
if min(lenx,leny) >= 4:
ar += [ [ lenx, leny] ]
if len(ar) > 0:
lenx, leny = np.mean ( ar, axis=0 )
lowest_len = min (lenx, leny)
#if debug:
# io.log_info ("lenx/leny:(%d/%d) " % (lenx, leny ) )
# io.log_info ("lowest_len = %f" % (lowest_len) )
if cfg.erode_mask_modifier != 0:
ero = int( lowest_len * ( 0.126 - lowest_len * 0.00004551365 ) * 0.01*cfg.erode_mask_modifier )
#if debug:
# io.log_info ("erode_size = %d" % (ero) )
if ero > 0:
img_face_mask_aaa = cv2.erode(img_face_mask_aaa, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(ero,ero)), iterations = 1 )
elif ero < 0:
img_face_mask_aaa = cv2.dilate(img_face_mask_aaa, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(-ero,-ero)), iterations = 1 )
if cfg.clip_hborder_mask_per > 0: #clip hborder before blur
prd_hborder_rect_mask_a = np.ones ( prd_face_mask_a.shape, dtype=np.float32)
prd_border_size = int ( prd_hborder_rect_mask_a.shape[1] * cfg.clip_hborder_mask_per )
prd_hborder_rect_mask_a[:,0:prd_border_size,:] = 0
prd_hborder_rect_mask_a[:,-prd_border_size:,:] = 0
prd_hborder_rect_mask_a[-prd_border_size:,:,:] = 0
prd_hborder_rect_mask_a = np.expand_dims(cv2.blur(prd_hborder_rect_mask_a, (prd_border_size, prd_border_size) ),-1)
img_prd_hborder_rect_mask_a = cv2.warpAffine( prd_hborder_rect_mask_a, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC )
img_prd_hborder_rect_mask_a = np.expand_dims (img_prd_hborder_rect_mask_a, -1)
img_face_mask_aaa *= img_prd_hborder_rect_mask_a
img_face_mask_aaa = np.clip( img_face_mask_aaa, 0, 1.0 )
#if debug:
# debugs += [img_face_mask_aaa.copy()]
if cfg.blur_mask_modifier > 0:
blur = int( lowest_len * 0.10 * 0.01*cfg.blur_mask_modifier )
#if debug:
# io.log_info ("blur_size = %d" % (blur) )
if blur > 0:
img_face_mask_aaa = cv2.blur(img_face_mask_aaa, (blur, blur) )
img_face_mask_aaa = np.clip( img_face_mask_aaa, 0, 1.0 )
#if debug:
# debugs += [img_face_mask_aaa.copy()]
if 'seamless' not in cfg.mode and cfg.color_transfer_mode != 0:
if cfg.color_transfer_mode == 1:
#if debug:
# debugs += [ np.clip( cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
prd_face_bgr = imagelib.reinhard_color_transfer ( np.clip( (prd_face_bgr*255).astype(np.uint8), 0, 255),
np.clip( (dst_face_bgr*255).astype(np.uint8), 0, 255),
source_mask=prd_face_mask_a, target_mask=prd_face_mask_a)
prd_face_bgr = np.clip( prd_face_bgr.astype(np.float32) / 255.0, 0.0, 1.0)
#if debug:
# debugs += [ np.clip( cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
elif cfg.color_transfer_mode == 2:
#if debug:
# debugs += [ np.clip( cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
prd_face_bgr = imagelib.linear_color_transfer (prd_face_bgr, dst_face_bgr)
prd_face_bgr = np.clip( prd_face_bgr, 0.0, 1.0)
#if debug:
# debugs += [ np.clip( cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
if cfg.mode == 'hist-match-bw':
prd_face_bgr = cv2.cvtColor(prd_face_bgr, cv2.COLOR_BGR2GRAY)
prd_face_bgr = np.repeat( np.expand_dims (prd_face_bgr, -1), (3,), -1 )
if cfg.mode == 'hist-match' or cfg.mode == 'hist-match-bw':
#if debug:
# debugs += [ cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ) ]
hist_mask_a = np.ones ( prd_face_bgr.shape[:2] + (1,) , dtype=np.float32)
if cfg.masked_hist_match:
hist_mask_a *= prd_face_mask_a
white = (1.0-hist_mask_a)* np.ones ( prd_face_bgr.shape[:2] + (1,) , dtype=np.float32)
hist_match_1 = prd_face_bgr*hist_mask_a + white
hist_match_1[ hist_match_1 > 1.0 ] = 1.0
hist_match_2 = dst_face_bgr*hist_mask_a + white
hist_match_2[ hist_match_1 > 1.0 ] = 1.0
prd_face_bgr = imagelib.color_hist_match(hist_match_1, hist_match_2, cfg.hist_match_threshold )
#if cfg.masked_hist_match:
# prd_face_bgr -= white
if cfg.mode == 'hist-match-bw':
prd_face_bgr = prd_face_bgr.astype(dtype=np.float32)
out_img = cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, out_img, cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT )
out_img = np.clip(out_img, 0.0, 1.0)
#if debug:
# debugs += [out_img.copy()]
if cfg.mode == 'overlay':
pass
if 'seamless' in cfg.mode:
#mask used for cv2.seamlessClone
img_face_seamless_mask_a = None
img_face_mask_a = img_face_mask_aaa[...,0:1]
for i in range(1,10):
a = img_face_mask_a > i / 10.0
if len(np.argwhere(a)) == 0:
continue
img_face_seamless_mask_a = img_face_mask_aaa[...,0:1].copy()
img_face_seamless_mask_a[a] = 1.0
img_face_seamless_mask_a[img_face_seamless_mask_a <= i / 10.0] = 0.0
break
try:
#calc same bounding rect and center point as in cv2.seamlessClone to prevent jittering (not flickering)
l,t,w,h = cv2.boundingRect( (img_face_seamless_mask_a*255).astype(np.uint8) )
s_maskx, s_masky = int(l+w/2), int(t+h/2)
out_img = cv2.seamlessClone( (out_img*255).astype(np.uint8), img_bgr_uint8, (img_face_seamless_mask_a*255).astype(np.uint8), (s_maskx,s_masky) , cv2.NORMAL_CLONE )
out_img = out_img.astype(dtype=np.float32) / 255.0
except Exception as e:
#seamlessClone may fail in some cases
e_str = traceback.format_exc()
if 'MemoryError' in e_str:
raise Exception("Seamless fail: " + e_str) #reraise MemoryError in order to reprocess this data by other processes
else:
print ("Seamless fail: " + e_str)
#if debug:
# debugs += [out_img.copy()]
out_img = img_bgr*(1-img_face_mask_aaa) + (out_img*img_face_mask_aaa)
if 'seamless' in cfg.mode and cfg.color_transfer_mode != 0:
out_face_bgr = cv2.warpAffine( out_img, face_mat, (output_size, output_size) )
if cfg.color_transfer_mode == 1:
#if debug:
# debugs += [ np.clip( cv2.warpAffine( out_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
face_mask_aaa = cv2.warpAffine( img_face_mask_aaa, face_mat, (output_size, output_size) )
new_out_face_bgr = imagelib.reinhard_color_transfer ( np.clip( (out_face_bgr*255).astype(np.uint8), 0, 255),
np.clip( (dst_face_bgr*255).astype(np.uint8), 0, 255),
source_mask=face_mask_aaa, target_mask=face_mask_aaa)
new_out_face_bgr = np.clip( new_out_face_bgr.astype(np.float32) / 255.0, 0.0, 1.0)
#if debug:
# debugs += [ np.clip( cv2.warpAffine( new_out_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
elif cfg.color_transfer_mode == 2:
#if debug:
# debugs += [ np.clip( cv2.warpAffine( out_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
new_out_face_bgr = imagelib.linear_color_transfer (out_face_bgr, dst_face_bgr)
new_out_face_bgr = np.clip( new_out_face_bgr, 0.0, 1.0)
#if debug:
# debugs += [ np.clip( cv2.warpAffine( new_out_face_bgr, face_output_mat, img_size, np.zeros(img_bgr.shape, dtype=np.float32), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT ), 0, 1.0) ]
new_out = cv2.warpAffine( new_out_face_bgr, face_mat, img_size, img_bgr.copy(), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT )
out_img = np.clip( img_bgr*(1-img_face_mask_aaa) + (new_out*img_face_mask_aaa) , 0, 1.0 )
if cfg.mode == 'seamless-hist-match':
out_face_bgr = cv2.warpAffine( out_img, face_mat, (output_size, output_size) )
new_out_face_bgr = imagelib.color_hist_match(out_face_bgr, dst_face_bgr, cfg.hist_match_threshold)
new_out = cv2.warpAffine( new_out_face_bgr, face_mat, img_size, img_bgr.copy(), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT )
out_img = np.clip( img_bgr*(1-img_face_mask_aaa) + (new_out*img_face_mask_aaa) , 0, 1.0 )
cfg_mp = cfg.motion_blur_power / 100.0
if cfg_mp != 0:
k_size = int(frame_info.motion_power*cfg_mp)
if k_size >= 1:
k_size = np.clip (k_size+1, 2, 50)
if cfg.super_resolution:
k_size *= 2
out_face_bgr = cv2.warpAffine( out_img, face_mat, (output_size, output_size) )
new_out_face_bgr = imagelib.LinearMotionBlur (out_face_bgr, k_size , frame_info.motion_deg)
new_out = cv2.warpAffine( new_out_face_bgr, face_mat, img_size, img_bgr.copy(), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC, cv2.BORDER_TRANSPARENT )
out_img = np.clip( img_bgr*(1-img_face_mask_aaa) + (new_out*img_face_mask_aaa) , 0, 1.0 )
if cfg.color_degrade_power != 0:
#if debug:
# debugs += [out_img.copy()]
out_img_reduced = imagelib.reduce_colors(out_img, 256)
if cfg.color_degrade_power == 100:
out_img = out_img_reduced
else:
alpha = cfg.color_degrade_power / 100.0
out_img = (out_img*(1.0-alpha) + out_img_reduced*alpha)
if cfg.export_mask_alpha:
out_img = np.concatenate ( [out_img, img_face_mask_aaa[:,:,0:1]], -1 )
out_merging_mask = img_face_mask_aaa
#if debug:
# debugs += [out_img.copy()]
return out_img, out_merging_mask
def ConvertMasked (cfg, frame_info):
img_bgr_uint8 = cv2_imread(frame_info.filename)
img_bgr_uint8 = imagelib.normalize_channels (img_bgr_uint8, 3)
img_bgr = img_bgr_uint8.astype(np.float32) / 255.0
outs = []
for face_num, img_landmarks in enumerate( frame_info.landmarks_list ):
out_img, out_img_merging_mask = ConvertMaskedFace (cfg, frame_info, img_bgr_uint8, img_bgr, img_landmarks)
outs += [ (out_img, out_img_merging_mask) ]
#Combining multiple face outputs
final_img = None
for img, merging_mask in outs:
h,w,c = img.shape
if final_img is None:
final_img = img
else:
merging_mask = merging_mask[...,0:1]
if c == 3:
final_img = final_img*(1-merging_mask) + img*merging_mask
elif c == 4:
final_img_bgr = final_img[...,0:3]*(1-merging_mask) + img[...,0:3]*merging_mask
final_img_mask = np.clip ( final_img[...,3:4] + img[...,3:4], 0, 1 )
final_img = np.concatenate ( [final_img_bgr, final_img_mask], -1 )
return (final_img*255).astype(np.uint8)