Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,243 Bytes
26557da |
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 |
import os
import cv2
import requests
import torch
import numpy as np
import PIL.Image
import PIL.ImageOps
from insightface.app import FaceAnalysis
from facexlib.parsing import init_parsing_model
from torchvision.transforms.functional import normalize
from typing import Union, Optional
def _img2tensor(img: np.ndarray, bgr2rgb: bool = True) -> torch.Tensor:
if bgr2rgb:
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = img.astype(np.float32) / 255.0
img = np.transpose(img, (2, 0, 1))
return torch.from_numpy(img)
def _pad_to_square(img: np.ndarray, pad_color: int = 255) -> np.ndarray:
h, w, _ = img.shape
if h == w:
return img
if h > w:
pad_size = (h - w) // 2
padded_img = cv2.copyMakeBorder(
img,
0,
0,
pad_size,
h - w - pad_size,
cv2.BORDER_CONSTANT,
value=[pad_color] * 3,
)
else:
pad_size = (w - h) // 2
padded_img = cv2.copyMakeBorder(
img,
pad_size,
w - h - pad_size,
0,
0,
cv2.BORDER_CONSTANT,
value=[pad_color] * 3,
)
return padded_img
class FaceProcessor:
def __init__(self, antelopv2_path=".", device: Optional[torch.device] = None):
if device is None:
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
else:
self.device = device
providers = (
["CUDAExecutionProvider"]
if self.device.type == "cuda"
else ["CPUExecutionProvider"]
)
self.app = FaceAnalysis(
name="antelopev2", root=antelopv2_path, providers=providers
)
self.app.prepare(ctx_id=0, det_size=(640, 640))
self.parsing_model = init_parsing_model(
model_name="bisenet", device=self.device
)
self.parsing_model.eval()
print("FaceProcessor initialized successfully.")
def process(
self,
image: Union[str, PIL.Image.Image],
resize_to: int = 512,
border_thresh: int = 10,
face_crop_scale: float = 1.5,
extra_input: bool = False,
) -> PIL.Image.Image:
if isinstance(image, str):
if image.startswith("http://") or image.startswith("https://"):
image = PIL.Image.open(requests.get(image, stream=True, timeout=10).raw)
elif os.path.isfile(image):
image = PIL.Image.open(image)
else:
raise ValueError(
f"Input string is not a valid URL or file path: {image}"
)
elif not isinstance(image, PIL.Image.Image):
raise TypeError(
"Input must be a file path, a URL, or a PIL.Image.Image object."
)
image = PIL.ImageOps.exif_transpose(image).convert("RGB")
frame = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
faces = self.app.get(frame)
h, w, _ = frame.shape
image_to_process = None
if not faces:
print(
"[Warning] No face detected. Using the whole image, padded to square."
)
image_to_process = _pad_to_square(frame, pad_color=255)
else:
largest_face = max(
faces, key=lambda f: (f.bbox[2] - f.bbox[0]) * (f.bbox[3] - f.bbox[1])
)
x1, y1, x2, y2 = map(int, largest_face.bbox)
is_close_to_border = (
x1 <= border_thresh
and y1 <= border_thresh
and x2 >= w - border_thresh
and y2 >= h - border_thresh
)
if is_close_to_border:
print(
"[Info] Face is close to border, padding original image to square."
)
image_to_process = _pad_to_square(frame, pad_color=255)
else:
cx, cy = (x1 + x2) // 2, (y1 + y2) // 2
side = int(max(x2 - x1, y2 - y1) * face_crop_scale)
half = side // 2
left = max(cx - half, 0)
top = max(cy - half, 0)
right = min(cx + half, w)
bottom = min(cy + half, h)
cropped_face = frame[top:bottom, left:right]
image_to_process = _pad_to_square(cropped_face, pad_color=255)
image_resized = cv2.resize(
image_to_process, (resize_to, resize_to), interpolation=cv2.INTER_AREA
)
face_tensor = (
_img2tensor(image_resized, bgr2rgb=True).unsqueeze(0).to(self.device)
)
with torch.no_grad():
normalized_face = normalize(face_tensor, [0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
parsing_out = self.parsing_model(normalized_face)[0]
parsing_mask = parsing_out.argmax(dim=1, keepdim=True)
background_mask_np = (parsing_mask.squeeze().cpu().numpy() == 0).astype(
np.uint8
)
white_background = np.ones_like(image_resized, dtype=np.uint8) * 255
mask_3channel = cv2.cvtColor(background_mask_np * 255, cv2.COLOR_GRAY2BGR)
result_img_bgr = np.where(mask_3channel == 255, white_background, image_resized)
result_img_rgb = cv2.cvtColor(result_img_bgr, cv2.COLOR_BGR2RGB)
img_white_bg = PIL.Image.fromarray(result_img_rgb)
if extra_input:
# 2. Create image with transparent background (new logic)
# Create an alpha channel: 255 for foreground (not background), 0 for background
alpha_channel = (parsing_mask.squeeze().cpu().numpy() != 0).astype(
np.uint8
) * 255
# Convert the resized BGR image to RGB
image_resized_rgb = cv2.cvtColor(image_resized, cv2.COLOR_BGR2RGB)
# Stack RGB channels with the new alpha channel
rgba_image = np.dstack((image_resized_rgb, alpha_channel))
# Create PIL image from the RGBA numpy array
img_transparent_bg = PIL.Image.fromarray(rgba_image, "RGBA")
return img_white_bg, img_transparent_bg
else:
return img_white_bg
|