File size: 8,536 Bytes
bf44779 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 |
from hashlib import sha1
from pathlib import Path
import cv2
import gradio as gr
import numpy as np
from PIL import Image
import PIL
import torch
from torchvision import transforms
import torch.nn.functional as F
def estimate_foreground_ml(image, alpha, return_background=False):
"""
Estimates the foreground and background of an image based on an alpha mask.
Parameters:
- image: numpy array of shape (H, W, 3), the input RGB image.
- alpha: numpy array of shape (H, W), the alpha mask with values ranging from 0 to 1.
- return_background: boolean, if True, both foreground and background are returned.
Returns:
- If return_background is False, returns only the foreground.
- If return_background is True, returns a tuple (foreground, background).
"""
# Estimating foreground
# Expand alpha dimensions from (H, W) to (H, W, 1) to make it compatible for element-wise multiplication with the RGB image
foreground = image * alpha
if return_background:
# Estimating background
# Inverse alpha mask to isolate background
background_alpha = 1 - alpha
# Assuming a white background. This can be changed based on the application or estimated from the image.
background = (image * background_alpha) + (1 - background_alpha) * 255
return foreground, background
return foreground
def load_entire_model(taskname):
model_ls = []
if (taskname == "mask"):
model = torch.jit.load(Path("./models/sod.pt"))
model.eval()
model_ls.append(model)
elif (taskname == "matting"):
model = torch.jit.load(Path("./models/trimap.pt"))
model.eval()
model_ls.append(model)
model = torch.jit.load(Path("./models/matting.pt"))
model.eval()
model_ls.append(model)
else:
model_ls = []
return model_ls
model_names = [
"matting",
"mask"
]
model_dict = {
name: None
for name in model_names
}
last_result = {
"cache_key": None,
"algorithm": None,
}
def image_matting(
image: PIL.Image.Image,
result_type: str,
bg_color: str,
algorithm: str,
morph_op: str,
morph_op_factor: float,
) -> np.ndarray:
image_np = np.ascontiguousarray(image)
width, height = image_np.shape[1], image_np.shape[0]
cache_key = sha1(image_np).hexdigest()
if cache_key == last_result["cache_key"] and algorithm == last_result["algorithm"]:
alpha = last_result["alpha"]
else:
model = load_entire_model(algorithm)
transform = transforms.Compose([
# transforms.ToPILImage(),
transforms.Resize((798, 798)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
if (algorithm == "mask"):
input_tensor = transform(image).unsqueeze(0)
with torch.no_grad():
alpha = model[0](input_tensor).float()
alpha = F.interpolate(alpha, [height, width], mode="bilinear")
alpha = np.array(alpha* 255.).astype(np.uint8)[0][0]
alpha = np.stack((alpha,alpha,alpha),axis=2)
else:
transform2 = transforms.Compose([
transforms.Resize((800, 800)),
transforms.ToTensor(),
# transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
input_tensor = transform(image).unsqueeze(0)
with torch.no_grad():
output = model[0](input_tensor).float()
output = F.interpolate(output, [height, width], mode="bilinear")
trimap = np.array(output[0][0])
ratio = 0.05
site = np.where(trimap > 0)
try:
bbox = [np.min(site[1]), np.min(site[0]), np.max(site[1]), np.max(site[0])]
except:
bbox = [0, 0, width, height]
x0, y0, x1, y1 = bbox
H = y1 - y0
W = x1 - x0
x0 = int(max(0, x0 - ratio * W))
x1 = int(min(width, x1 + ratio * W) )
y0 = int(max(0, y0 - ratio * H) )
y1 = int(min(height, y1 + ratio * H) )
Image_input = image.crop((x0, y0, x1, y1))
# Image_input.save('image.png')
input_tensor = transform2(Image_input).unsqueeze(0)
trimap = trimap[y0:y1, x0:x1]
trimap = np.where(trimap < 1, 0, trimap)
trimap = np.where(trimap > 1, 255, trimap)
trimap = np.where(trimap == 1, 128, trimap)
# cv2.imwrite("trimap.png", trimap)
trimap = Image.fromarray(np.uint8(trimap)).convert('L')
input_tensor2 = transform2(trimap).unsqueeze(0)
with torch.no_grad():
output = model[1]({'image': input_tensor, 'trimap': input_tensor2})['phas']
output = F.interpolate(output, [Image_input.size[1],Image_input.size[0]], mode="bilinear")[0].numpy()
numpy_image = (output * 255.).astype(np.uint8) # Scale to [0, 255] and convert to uint8
# Step 4: Remove the channel dimension since it's a grayscale image
numpy_image = numpy_image.squeeze(0) # Convert from (1, H, W) to (H, W)
pil_image = Image.fromarray(numpy_image, mode='L')
alpha = Image.new(mode='RGB', size=image.size)
alpha.paste(pil_image, (x0, y0))
# alpha.save('tmp.png')
alpha = np.array(alpha).astype(np.uint8)
last_result["cache_key"] = cache_key
last_result["algorithm"] = algorithm
last_result["alpha"] = alpha
# alpha = (alpha * 255).astype(np.uint8)
image = np.array(image)
kernel = np.ones((morph_op_factor, morph_op_factor), np.uint8)
if morph_op == "Dilate":
alpha = cv2.dilate(alpha, kernel, iterations=int(morph_op_factor))
elif morph_op == "Erode":
alpha = cv2.erode(alpha, kernel, iterations=int(morph_op_factor))
else:
alpha = alpha
alpha = (alpha / 255).astype("float32")
image = (image / 255.0).astype("float32")
fg = estimate_foreground_ml(image, alpha)
if result_type == "Remove BG":
result = fg
elif result_type == "Replace BG":
bg_r = int(bg_color[1:3], base=16)
bg_g = int(bg_color[3:5], base=16)
bg_b = int(bg_color[5:7], base=16)
bg = np.zeros_like(fg)
bg[:, :, 0] = bg_r / 255.
bg[:, :, 1] = bg_g / 255.
bg[:, :, 2] = bg_b / 255.
result = alpha * image + (1 - alpha) * bg
result = np.clip(result, 0, 1)
else:
result = alpha
return result
def main():
with gr.Blocks() as app:
gr.Markdown("Salient Object Matting")
with gr.Row(variant="panel"):
image_input = gr.Image(type='pil')
image_output = gr.Image()
with gr.Row(variant="panel"):
result_type = gr.Radio(
label="Mode",
show_label=True,
choices=[
"Remove BG",
"Replace BG",
"Generate Mask",
],
value="Remove BG",
)
bg_color = gr.ColorPicker(
label="BG Color",
show_label=True,
value="#000000",
)
algorithm = gr.Dropdown(
label="Algorithm",
show_label=True,
choices=model_names,
value="matting"
)
with gr.Row(variant="panel"):
morph_op = gr.Radio(
label="Post-process",
show_label=True,
choices=[
"None",
"Erode",
"Dilate",
],
value="None",
)
morph_op_factor = gr.Slider(
label="Factor",
show_label=True,
minimum=3,
maximum=20,
value=3,
step=2,
)
run_button = gr.Button("Run")
run_button.click(
image_matting,
inputs=[
image_input,
result_type,
bg_color,
algorithm,
morph_op,
morph_op_factor,
],
outputs=image_output,
)
app.launch()
if __name__ == "__main__":
main()
|