File size: 9,606 Bytes
2fe55e2 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 |
import abc
from typing import Optional
import cv2
import torch
import numpy as np
from loguru import logger
from lama_cleaner.helper import (
boxes_from_mask,
resize_max_size,
pad_img_to_modulo,
switch_mps_device,
)
from lama_cleaner.schema import Config, HDStrategy
class InpaintModel:
name = "base"
min_size: Optional[int] = None
pad_mod = 8
pad_to_square = False
def __init__(self, device, **kwargs):
"""
Args:
device:
"""
device = switch_mps_device(self.name, device)
self.device = device
self.init_model(device, **kwargs)
@abc.abstractmethod
def init_model(self, device, **kwargs):
...
@staticmethod
@abc.abstractmethod
def is_downloaded() -> bool:
...
@abc.abstractmethod
def forward(self, image, mask, config: Config):
"""Input images and output images have same size
images: [H, W, C] RGB
masks: [H, W, 1] 255 为 masks 区域
return: BGR IMAGE
"""
...
def _pad_forward(self, image, mask, config: Config):
origin_height, origin_width = image.shape[:2]
pad_image = pad_img_to_modulo(
image, mod=self.pad_mod, square=self.pad_to_square, min_size=self.min_size
)
pad_mask = pad_img_to_modulo(
mask, mod=self.pad_mod, square=self.pad_to_square, min_size=self.min_size
)
logger.info(f"final forward pad size: {pad_image.shape}")
result = self.forward(pad_image, pad_mask, config)
result = result[0:origin_height, 0:origin_width, :]
result, image, mask = self.forward_post_process(result, image, mask, config)
mask = mask[:, :, np.newaxis]
result = result * (mask / 255) + image[:, :, ::-1] * (1 - (mask / 255))
return result
def forward_post_process(self, result, image, mask, config):
return result, image, mask
@torch.no_grad()
def __call__(self, image, mask, config: Config):
"""
images: [H, W, C] RGB, not normalized
masks: [H, W]
return: BGR IMAGE
"""
inpaint_result = None
logger.info(f"hd_strategy: {config.hd_strategy}")
if config.hd_strategy == HDStrategy.CROP:
if max(image.shape) > config.hd_strategy_crop_trigger_size:
logger.info(f"Run crop strategy")
boxes = boxes_from_mask(mask)
crop_result = []
for box in boxes:
crop_image, crop_box = self._run_box(image, mask, box, config)
crop_result.append((crop_image, crop_box))
inpaint_result = image[:, :, ::-1]
for crop_image, crop_box in crop_result:
x1, y1, x2, y2 = crop_box
inpaint_result[y1:y2, x1:x2, :] = crop_image
elif config.hd_strategy == HDStrategy.RESIZE:
if max(image.shape) > config.hd_strategy_resize_limit:
origin_size = image.shape[:2]
downsize_image = resize_max_size(
image, size_limit=config.hd_strategy_resize_limit
)
downsize_mask = resize_max_size(
mask, size_limit=config.hd_strategy_resize_limit
)
logger.info(
f"Run resize strategy, origin size: {image.shape} forward size: {downsize_image.shape}"
)
inpaint_result = self._pad_forward(
downsize_image, downsize_mask, config
)
# only paste masked area result
inpaint_result = cv2.resize(
inpaint_result,
(origin_size[1], origin_size[0]),
interpolation=cv2.INTER_CUBIC,
)
original_pixel_indices = mask < 127
inpaint_result[original_pixel_indices] = image[:, :, ::-1][
original_pixel_indices
]
if inpaint_result is None:
inpaint_result = self._pad_forward(image, mask, config)
return inpaint_result
def _crop_box(self, image, mask, box, config: Config):
"""
Args:
image: [H, W, C] RGB
mask: [H, W, 1]
box: [left,top,right,bottom]
Returns:
BGR IMAGE, (l, r, r, b)
"""
box_h = box[3] - box[1]
box_w = box[2] - box[0]
cx = (box[0] + box[2]) // 2
cy = (box[1] + box[3]) // 2
img_h, img_w = image.shape[:2]
w = box_w + config.hd_strategy_crop_margin * 2
h = box_h + config.hd_strategy_crop_margin * 2
_l = cx - w // 2
_r = cx + w // 2
_t = cy - h // 2
_b = cy + h // 2
l = max(_l, 0)
r = min(_r, img_w)
t = max(_t, 0)
b = min(_b, img_h)
# try to get more context when crop around image edge
if _l < 0:
r += abs(_l)
if _r > img_w:
l -= _r - img_w
if _t < 0:
b += abs(_t)
if _b > img_h:
t -= _b - img_h
l = max(l, 0)
r = min(r, img_w)
t = max(t, 0)
b = min(b, img_h)
crop_img = image[t:b, l:r, :]
crop_mask = mask[t:b, l:r]
logger.info(f"box size: ({box_h},{box_w}) crop size: {crop_img.shape}")
return crop_img, crop_mask, [l, t, r, b]
def _calculate_cdf(self, histogram):
cdf = histogram.cumsum()
normalized_cdf = cdf / float(cdf.max())
return normalized_cdf
def _calculate_lookup(self, source_cdf, reference_cdf):
lookup_table = np.zeros(256)
lookup_val = 0
for source_index, source_val in enumerate(source_cdf):
for reference_index, reference_val in enumerate(reference_cdf):
if reference_val >= source_val:
lookup_val = reference_index
break
lookup_table[source_index] = lookup_val
return lookup_table
def _match_histograms(self, source, reference, mask):
transformed_channels = []
for channel in range(source.shape[-1]):
source_channel = source[:, :, channel]
reference_channel = reference[:, :, channel]
# only calculate histograms for non-masked parts
source_histogram, _ = np.histogram(source_channel[mask == 0], 256, [0, 256])
reference_histogram, _ = np.histogram(
reference_channel[mask == 0], 256, [0, 256]
)
source_cdf = self._calculate_cdf(source_histogram)
reference_cdf = self._calculate_cdf(reference_histogram)
lookup = self._calculate_lookup(source_cdf, reference_cdf)
transformed_channels.append(cv2.LUT(source_channel, lookup))
result = cv2.merge(transformed_channels)
result = cv2.convertScaleAbs(result)
return result
def _apply_cropper(self, image, mask, config: Config):
img_h, img_w = image.shape[:2]
l, t, w, h = (
config.croper_x,
config.croper_y,
config.croper_width,
config.croper_height,
)
r = l + w
b = t + h
l = max(l, 0)
r = min(r, img_w)
t = max(t, 0)
b = min(b, img_h)
crop_img = image[t:b, l:r, :]
crop_mask = mask[t:b, l:r]
return crop_img, crop_mask, (l, t, r, b)
def _run_box(self, image, mask, box, config: Config):
"""
Args:
image: [H, W, C] RGB
mask: [H, W, 1]
box: [left,top,right,bottom]
Returns:
BGR IMAGE
"""
crop_img, crop_mask, [l, t, r, b] = self._crop_box(image, mask, box, config)
return self._pad_forward(crop_img, crop_mask, config), [l, t, r, b]
class DiffusionInpaintModel(InpaintModel):
@torch.no_grad()
def __call__(self, image, mask, config: Config):
"""
images: [H, W, C] RGB, not normalized
masks: [H, W]
return: BGR IMAGE
"""
# boxes = boxes_from_mask(mask)
if config.use_croper:
crop_img, crop_mask, (l, t, r, b) = self._apply_cropper(image, mask, config)
crop_image = self._scaled_pad_forward(crop_img, crop_mask, config)
inpaint_result = image[:, :, ::-1]
inpaint_result[t:b, l:r, :] = crop_image
else:
inpaint_result = self._scaled_pad_forward(image, mask, config)
return inpaint_result
def _scaled_pad_forward(self, image, mask, config: Config):
longer_side_length = int(config.sd_scale * max(image.shape[:2]))
origin_size = image.shape[:2]
downsize_image = resize_max_size(image, size_limit=longer_side_length)
downsize_mask = resize_max_size(mask, size_limit=longer_side_length)
if config.sd_scale != 1:
logger.info(
f"Resize image to do sd inpainting: {image.shape} -> {downsize_image.shape}"
)
inpaint_result = self._pad_forward(downsize_image, downsize_mask, config)
# only paste masked area result
inpaint_result = cv2.resize(
inpaint_result,
(origin_size[1], origin_size[0]),
interpolation=cv2.INTER_CUBIC,
)
original_pixel_indices = mask < 127
inpaint_result[original_pixel_indices] = image[:, :, ::-1][
original_pixel_indices
]
return inpaint_result
|