Spaces:
Sleeping
Sleeping
Update rembg/bg.py
Browse files- rembg/bg.py +180 -35
rembg/bg.py
CHANGED
@@ -1,26 +1,142 @@
|
|
1 |
import io
|
2 |
-
from typing import Any, Union
|
3 |
-
import numpy as np
|
4 |
-
from PIL import Image
|
5 |
from enum import Enum
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
-
from rembg import new_session, remove
|
8 |
-
from rembg.sessions.base import BaseSession
|
9 |
-
from rembg.util.util import fix_image_orientation
|
10 |
|
11 |
class ReturnType(Enum):
|
12 |
BYTES = 0
|
13 |
PILLOW = 1
|
14 |
NDARRAY = 2
|
15 |
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
return_type = ReturnType.PILLOW
|
25 |
img = data
|
26 |
elif isinstance(data, bytes):
|
@@ -30,27 +146,56 @@ def remove_background(
|
|
30 |
return_type = ReturnType.NDARRAY
|
31 |
img = Image.fromarray(data)
|
32 |
else:
|
33 |
-
raise ValueError("Input type {} is not supported.".format(type(data))
|
34 |
|
|
|
35 |
img = fix_image_orientation(img)
|
36 |
-
session = new_session(model)
|
37 |
-
output = remove(
|
38 |
-
img,
|
39 |
-
alpha_matting=True,
|
40 |
-
alpha_matting_foreground_threshold=alpha_influence * 255,
|
41 |
-
alpha_matting_background_threshold=(1 - alpha_influence) * 255,
|
42 |
-
alpha_matting_erode_size=int(segmentation_strength * 20),
|
43 |
-
alpha_matting_smoothing=smoothing,
|
44 |
-
session=session
|
45 |
-
)
|
46 |
-
|
47 |
-
if return_type == ReturnType.PILLOW:
|
48 |
-
return output
|
49 |
-
elif return_type == ReturnType.NDARRAY:
|
50 |
-
return np.array(output)
|
51 |
-
else:
|
52 |
-
bio = io.BytesIO()
|
53 |
-
output.save(bio, "PNG")
|
54 |
-
bio.seek(0)
|
55 |
-
return bio.read()
|
56 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import io
|
|
|
|
|
|
|
2 |
from enum import Enum
|
3 |
+
from typing import Any, List, Optional, Tuple, Union
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
from cv2 import (
|
7 |
+
BORDER_DEFAULT,
|
8 |
+
MORPH_ELLIPSE,
|
9 |
+
MORPH_OPEN,
|
10 |
+
GaussianBlur,
|
11 |
+
getStructuringElement,
|
12 |
+
morphologyEx,
|
13 |
+
)
|
14 |
+
from PIL import Image, ImageOps
|
15 |
+
from PIL.Image import Image as PILImage
|
16 |
+
from pymatting.alpha.estimate_alpha_cf import estimate_alpha_cf
|
17 |
+
from pymatting.foreground.estimate_foreground_ml import estimate_foreground_ml
|
18 |
+
from pymatting.util.util import stack_images
|
19 |
+
from scipy.ndimage import binary_erosion
|
20 |
+
|
21 |
+
from .session_factory import new_session
|
22 |
+
from .sessions import sessions_class
|
23 |
+
from .sessions.base import BaseSession
|
24 |
+
|
25 |
+
kernel = getStructuringElement(MORPH_ELLIPSE, (3, 3))
|
26 |
|
|
|
|
|
|
|
27 |
|
28 |
class ReturnType(Enum):
|
29 |
BYTES = 0
|
30 |
PILLOW = 1
|
31 |
NDARRAY = 2
|
32 |
|
33 |
+
|
34 |
+
def alpha_matting_cutout(
|
35 |
+
img: PILImage,
|
36 |
+
mask: PILImage,
|
37 |
+
foreground_threshold: int,
|
38 |
+
background_threshold: int,
|
39 |
+
erode_structure_size: int,
|
40 |
+
) -> PILImage:
|
41 |
+
if img.mode == "RGBA" or img.mode == "CMYK":
|
42 |
+
img = img.convert("RGB")
|
43 |
+
|
44 |
+
img = np.asarray(img)
|
45 |
+
mask = np.asarray(mask)
|
46 |
+
|
47 |
+
is_foreground = mask > foreground_threshold
|
48 |
+
is_background = mask < background_threshold
|
49 |
+
|
50 |
+
structure = None
|
51 |
+
if erode_structure_size > 0:
|
52 |
+
structure = np.ones(
|
53 |
+
(erode_structure_size, erode_structure_size), dtype=np.uint8
|
54 |
+
)
|
55 |
+
|
56 |
+
is_foreground = binary_erosion(is_foreground, structure=structure)
|
57 |
+
is_background = binary_erosion(is_background, structure=structure, border_value=1)
|
58 |
+
|
59 |
+
trimap = np.full(mask.shape, dtype=np.uint8, fill_value=128)
|
60 |
+
trimap[is_foreground] = 255
|
61 |
+
trimap[is_background] = 0
|
62 |
+
|
63 |
+
img_normalized = img / 255.0
|
64 |
+
trimap_normalized = trimap / 255.0
|
65 |
+
|
66 |
+
alpha = estimate_alpha_cf(img_normalized, trimap_normalized)
|
67 |
+
foreground = estimate_foreground_ml(img_normalized, alpha)
|
68 |
+
cutout = stack_images(foreground, alpha)
|
69 |
+
|
70 |
+
cutout = np.clip(cutout * 255, 0, 255).astype(np.uint8)
|
71 |
+
cutout = Image.fromarray(cutout)
|
72 |
+
|
73 |
+
return cutout
|
74 |
+
|
75 |
+
|
76 |
+
def naive_cutout(img: PILImage, mask: PILImage) -> PILImage:
|
77 |
+
empty = Image.new("RGBA", (img.size), 0)
|
78 |
+
cutout = Image.composite(img, empty, mask)
|
79 |
+
return cutout
|
80 |
+
|
81 |
+
|
82 |
+
def get_concat_v_multi(imgs: List[PILImage]) -> PILImage:
|
83 |
+
pivot = imgs.pop(0)
|
84 |
+
for im in imgs:
|
85 |
+
pivot = get_concat_v(pivot, im)
|
86 |
+
return pivot
|
87 |
+
|
88 |
+
|
89 |
+
def get_concat_v(img1: PILImage, img2: PILImage) -> PILImage:
|
90 |
+
dst = Image.new("RGBA", (img1.width, img1.height + img2.height))
|
91 |
+
dst.paste(img1, (0, 0))
|
92 |
+
dst.paste(img2, (0, img1.height))
|
93 |
+
return dst
|
94 |
+
|
95 |
+
|
96 |
+
def post_process(mask: np.ndarray) -> np.ndarray:
|
97 |
+
"""
|
98 |
+
Post Process the mask for a smooth boundary by applying Morphological Operations
|
99 |
+
Research based on paper: https://www.sciencedirect.com/science/article/pii/S2352914821000757
|
100 |
+
args:
|
101 |
+
mask: Binary Numpy Mask
|
102 |
+
"""
|
103 |
+
mask = morphologyEx(mask, MORPH_OPEN, kernel)
|
104 |
+
mask = GaussianBlur(mask, (5, 5), sigmaX=2, sigmaY=2, borderType=BORDER_DEFAULT)
|
105 |
+
mask = np.where(mask < 127, 0, 255).astype(np.uint8) # convert again to binary
|
106 |
+
return mask
|
107 |
+
|
108 |
+
|
109 |
+
def apply_background_color(img: PILImage, color: Tuple[int, int, int, int]) -> PILImage:
|
110 |
+
r, g, b, a = color
|
111 |
+
colored_image = Image.new("RGBA", img.size, (r, g, b, a))
|
112 |
+
colored_image.paste(img, mask=img)
|
113 |
+
|
114 |
+
return colored_image
|
115 |
+
|
116 |
+
|
117 |
+
def fix_image_orientation(img: PILImage) -> PILImage:
|
118 |
+
return ImageOps.exif_transpose(img)
|
119 |
+
|
120 |
+
|
121 |
+
def download_models() -> None:
|
122 |
+
for session in sessions_class:
|
123 |
+
session.download_models()
|
124 |
+
|
125 |
+
|
126 |
+
def remove(
|
127 |
+
data: Union[bytes, PILImage, np.ndarray],
|
128 |
+
alpha_matting: bool = False,
|
129 |
+
alpha_matting_foreground_threshold: int = 240,
|
130 |
+
alpha_matting_background_threshold: int = 10,
|
131 |
+
alpha_matting_erode_size: int = 10,
|
132 |
+
session: Optional[BaseSession] = None,
|
133 |
+
only_mask: bool = False,
|
134 |
+
post_process_mask: bool = False,
|
135 |
+
bgcolor: Optional[Tuple[int, int, int, int]] = None,
|
136 |
+
*args: Optional[Any],
|
137 |
+
**kwargs: Optional[Any]
|
138 |
+
) -> Union[bytes, PILImage, np.ndarray]:
|
139 |
+
if isinstance(data, PILImage):
|
140 |
return_type = ReturnType.PILLOW
|
141 |
img = data
|
142 |
elif isinstance(data, bytes):
|
|
|
146 |
return_type = ReturnType.NDARRAY
|
147 |
img = Image.fromarray(data)
|
148 |
else:
|
149 |
+
raise ValueError("Input type {} is not supported.".format(type(data)))
|
150 |
|
151 |
+
# Fix image orientation
|
152 |
img = fix_image_orientation(img)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
|
154 |
+
if session is None:
|
155 |
+
session = new_session("u2net", *args, **kwargs)
|
156 |
+
|
157 |
+
masks = session.predict(img, *args, **kwargs)
|
158 |
+
cutouts = []
|
159 |
+
|
160 |
+
for mask in masks:
|
161 |
+
if post_process_mask:
|
162 |
+
mask = Image.fromarray(post_process(np.array(mask)))
|
163 |
+
|
164 |
+
if only_mask:
|
165 |
+
cutout = mask
|
166 |
+
|
167 |
+
elif alpha_matting:
|
168 |
+
try:
|
169 |
+
cutout = alpha_matting_cutout(
|
170 |
+
img,
|
171 |
+
mask,
|
172 |
+
alpha_matting_foreground_threshold,
|
173 |
+
alpha_matting_background_threshold,
|
174 |
+
alpha_matting_erode_size,
|
175 |
+
)
|
176 |
+
except ValueError:
|
177 |
+
cutout = naive_cutout(img, mask)
|
178 |
+
|
179 |
+
else:
|
180 |
+
cutout = naive_cutout(img, mask)
|
181 |
+
|
182 |
+
cutouts.append(cutout)
|
183 |
+
|
184 |
+
cutout = img
|
185 |
+
if len(cutouts) > 0:
|
186 |
+
cutout = get_concat_v_multi(cutouts)
|
187 |
+
|
188 |
+
if bgcolor is not None and not only_mask:
|
189 |
+
cutout = apply_background_color(cutout, bgcolor)
|
190 |
+
|
191 |
+
if ReturnType.PILLOW == return_type:
|
192 |
+
return cutout
|
193 |
+
|
194 |
+
if ReturnType.NDARRAY == return_type:
|
195 |
+
return np.asarray(cutout)
|
196 |
+
|
197 |
+
bio = io.BytesIO()
|
198 |
+
cutout.save(bio, "PNG")
|
199 |
+
bio.seek(0)
|
200 |
+
|
201 |
+
return bio.read()
|