def apply_fixes(self, original, face, mask, landmarks):
""" Apply fixes """
#TODO copies aren't likey neccesary and will slow calc... used when isolating issues
new_image = face[:, :, :3].copy()
image_mask = mask[:, :, :3].copy()
frame = original[:, :, :3].copy()
#TODO - force default for args.sharpen_image to ensure it isn't None
if self.args.sharpen_image is not None and self.args.sharpen_image.lower(
) != "none":
new_image = self.sharpen(new_image, self.args.sharpen_image)
if self.args.avg_color_adjust:
new_image = self.color_adjust(new_image, frame, image_mask)
if self.args.match_histogram:
new_image = self.color_hist_match(new_image, frame, image_mask)
if self.args.seamless_clone:
blended = self.seamless_clone(new_image, frame, image_mask)
else:
foreground = new_image * image_mask
background = frame * (1.0 - image_mask)
blended = foreground + background
np.clip(blended, 0.0, 255.0, out=blended)
if self.args.draw_transparent:
# Adding a 4th channel should happen after all other channel operations
# Add mask as 4th channel for saving as alpha on supported output formats
blended = dfl_full(landmarks, blended, channels=4)
return np.rint(blended).astype('uint8')
def dfl(self, **kwargs):
""" DFaker Mask """
logger.trace("Getting mask")
dummy = np.zeros((kwargs["image_size"][1], kwargs["image_size"][0], 3), dtype='float32')
mask = dfl_full(kwargs["landmarks"], dummy, channels=3)
mask = self.intersect_rect(mask, **kwargs)
return mask
def patch_image(self, image, detected_face):
""" Patch the image """
logger.trace("Patching image")
image = image.astype('float32')
image_size = (image.shape[1], image.shape[0])
coverage = int(self.training_coverage_ratio * self.training_size)
padding = (self.training_size - coverage) // 2
logger.trace("coverage: %s, padding: %s", coverage, padding)
self.crop = slice(padding, self.training_size - padding)
if not self.mask: # Init the mask on first image
self.mask = Mask(self.args.mask_type, self.training_size, padding,
self.crop)
detected_face.load_aligned(image,
size=self.training_size,
align_eyes=False)
new_image = self.get_new_image(image, detected_face, coverage,
image_size)
image_mask = self.get_image_mask(detected_face, image_size)
if self.args.draw_transparent:
new_image = dfl_full(
detected_face.landmarks_as_xy, new_image, channels=4
) #Add mask as 4th channel for saving as alpha on supported output formats
#This make sure that all the arrays match in size for later actions despite not actually using alpha in any way.
image_mask = cv2.cvtColor(image_mask, cv2.COLOR_RGB2RGBA)
image = cv2.cvtColor(image, cv2.COLOR_RGB2RGBA)
patched_face = self.apply_fixes(image, new_image, image_mask,
image_size)
logger.trace("Patched image")
return patched_face
def apply_fixes(self, original, face, mask, image_size, landmarks):
""" Apply fixes """
new_image = face[:, :, :3].copy()
image_mask = mask[:, :, :3].copy()
frame = original[:, :, :3].copy()
if self.args.sharpen_image is not None and self.args.sharpen_image.lower(
) != "none":
np.clip(new_image, 0.0, 255.0, out=new_image)
if self.args.sharpen_image == "box_filter":
kernel = np.ones((3, 3)) * (-1)
kernel[1, 1] = 9
new_image = cv2.filter2D(new_image, -1, kernel) # pylint: disable=no-member
elif self.args.sharpen_image == "gaussian_filter":
blur = cv2.GaussianBlur(new_image, (0, 0), 3.0) # pylint: disable=no-member
new_image = cv2.addWeighted(
new_image, # pylint: disable=no-member
1.5,
blur,
-0.5,
0,
new_image)
if self.args.avg_color_adjust:
for _ in [0, 1]:
np.clip(new_image, 0.0, 255.0, out=new_image)
diff = frame - new_image
avg_diff = np.sum(diff * image_mask, axis=(0, 1))
adjustment = avg_diff / np.sum(image_mask, axis=(0, 1))
new_image = new_image + adjustment
if self.args.match_histogram:
np.clip(new_image, 0.0, 255.0, out=new_image)
new_image = self.color_hist_match(new_image, frame, image_mask)
if self.args.seamless_clone:
h, w, _ = frame.shape
h = h // 2
w = w // 2
y_indices, x_indices, _ = np.nonzero(image_mask)
y_crop = slice(np.min(y_indices), np.max(y_indices))
x_crop = slice(np.min(x_indices), np.max(x_indices))
y_center = int(
np.rint((np.max(y_indices) + np.min(y_indices)) / 2) + h)
x_center = int(
np.rint((np.max(x_indices) + np.min(x_indices)) / 2) + w)
'''
# test with average of centroid rather than the h /2 , w/2 center
y_center = int(np.rint(np.average(y_indices) + h)
x_center = int(np.rint(np.average(x_indices) + w)
'''
insertion = np.rint(new_image[y_crop, x_crop, :]).astype('uint8')
insertion_mask = image_mask[y_crop, x_crop, :]
insertion_mask[insertion_mask != 0] = 255
insertion_mask = insertion_mask.astype('uint8')
prior = np.pad(frame, ((h, h), (w, w), (0, 0)),
'constant').astype('uint8')
blended = cv2.seamlessClone(
insertion, # pylint: disable=no-member
prior,
insertion_mask,
(x_center, y_center),
cv2.NORMAL_CLONE) # pylint: disable=no-member
blended = blended[h:-h, w:-w]
else:
foreground = new_image * image_mask
background = frame * (1.0 - image_mask)
blended = foreground + background
np.clip(blended, 0.0, 255.0, out=blended)
if self.args.draw_transparent:
# Adding a 4th channel should happen after all other channel operations
# Add mask as 4th channel for saving as alpha on supported output formats
blended = dfl_full(landmarks, blended, channels=4)
return np.rint(blended).astype('uint8')